MIT Technology Review Explains: Let our writers untangle the complex, messy world of science and technology to help you understand what’s coming next. You can read more from the series here.

Late last month, Anthropic released a report arguing that the model GLM-5.3 poses a major cyberattack risk. The report’s core concern is that the powerful new model, released in August by the Chinese company Z.ai, is freely available for anyone to download and modify. Anthropic’s assertion is only the latest move in a long-running debate about whether frontier AI models should be made open, like GLM-5.3, or should stay locked behind the doors of the frontier labs. 

These days, much of the debate about how to handle open models is playing out in public—and there’s a lot of money resting on its outcome. Nvidia CEO Jensen Huang, who has a lot to gain if people continue buying their own hardware to run open models, advocates enthusiastically for the United States to develop more of them, whereas Anthropic’s Dario Amodei, who runs a company that could be undermined by cheap, powerful models, advocates for a more cautious approach. But beneath all these commercial incentives, there is real substance to disagreements about open models.

“I think a lot of the way you come down on the openness debate is influenced by how you come down on the AI debate,” says David Gray Widder, an assistant professor at the University of Texas at Austin School of Information. Those who are concerned about AI safety warn that open models can easily be misused by bad actors, whereas those who focus more on AI’s practical potential urge that this power should not sit behind the closed doors of just a few big companies, and that open models could help spread it more widely.

There are a range of ways to handle open models between banning them entirely and releasing them willy-nilly. But any approach that a government or a society might take still involves trade-offs between safety and democratization. Deciding whether to support or limit the production of open models requires understanding those risks and benefits in depth.

What are the benefits of open models?

Let’s start with the basics. A closed model, like the models that power ChatGPT and Claude, can only be accessed through a website, app, or API. An open model, on the other hand, can be directly downloaded: The details about its internal structure, encoded in a long series of numbers known as the model weights, are publicly available. Such models are sometimes called “open-weight models,” to distinguish them from models for which even more information, such as details about the training procedure, is available. 

Right now, closed models are more capable than open models—as of publication, the top eight models on the Artificial Analysis LLM leaderboard are all closed, though open models are only a handful of months behind in development. But there are several reasons why someone might opt to use an open model rather than a closed one. Because open models are downloadable, people can run them on their own computers or hard drives. They can even retrain them to behave differently, if they have the expertise and resources to do so. 

If you’re an academic researcher working to understand and improve AI systems, you almost certainly need to work with open models in your research. When working with closed models, researchers can’t do much more than send prompts and observe the responses. But researchers can observe the inner workings of open models, as well as the “chain of thought” that the models use to think through a problem before answering. They can also combine multiple open models to build new systems, compress them to make them more efficient, and fine-tune them for new applications.

Businesses, too, have reasons to prefer open models. Even if you depend on an external provider to run them for you, open models still typically cost less than closed models, according to a recent report from Mozilla. And if businesses want to use AI to process sensitive data or proprietary information, running a local model on their own hardware may be the best way to keep their operations secure.

But perhaps the most important quality of open models is that they challenge the tight control the frontier labs have over how most of us use AI. “The clearest, biggest argument [for open models] is around preventing mass concentration of power—it is about making sure that this foundational technology is a public good,” says Elizabeth Seger, senior policy advisor at the Tony Blair Institute for Global Change. “I think we’re at this juncture where the question of who controls the infrastructure that allows AI to operate is as big of a question as who gets to control information on the internet.”

Seger sees the future of AI playing out in different ways, depending on the role that open models have. In one possible future, frontier AI systems are controlled entirely by select people, companies, or governments, and as those systems grow more powerful, those groups have increasing control over what our world looks like. The world’s most sophisticated hacking models, for example, are currently all made in the United States, so it’s entirely up to US companies and the US government whether other countries can use them to shore up their own cyber defenses. If those companies remain in the lead, we could in principle wind up in a world in which the United States is capable of hacking any other country, but no one is capable of hacking it.

Frontier open models could change that balance. “If every individual had a loyal-to-them AI system that was safe [and secure], I think that would be a much less concerning world than one in which they have to just rely on a government that says, ‘Yeah, we’ve got a super intelligent AI system, it’s got this plan, it’s inscrutable to all of you, just trust us,’” says Hamish Hobbs, director of AI policy at the Centre for Long-Term Resilience.

But open models are not a geopolitical panacea. Paul Azunre, who cofounded Ghana NLP, an organization that builds open-source AI tools for low-resource Ghanaian languages, notes that open models can also function as a way for companies to benefit from the labor and expertise of less-privileged nations and consolidate more power in their own hands. He identifies a pattern in which AI companies release models openly, observe how the community uses and modifies them, and then integrate those same improvements into their own products—without any credit or compensation for the people who originated those innovations. Mark Zuckerberg himself has said that the ability to integrate community improvements into Meta’s stack is a key benefit of releasing open models.

“I think the ideal of open source is a very noble one, and there’s a lot of value that it brings,” Azunre says. “But the way it’s actually implemented in practice, it tends to be a very extractive approach.”

Are open models riskier than closed ones?

The past few months have made it clear that safety remains a gnarly problem for closed models. There’s the autonomous hack that OpenAI agents executed against Hugging Face in July, but human misuse is a real threat too: According to a recent risk report from Anthropic, several extensive hacking operations have been undertaken with the assistance of Claude. 

So closed models are risky enough—and open models would likely be much easier to exploit. Before Anthropic passes a user’s query to one of its closed models, it uses automated systems to detect whether that query relates to hacking, biological weapons, or other risks. When it detects that users have outsmarted those filters, it can shut down their accounts. But when someone downloads a model and runs it on their own hardware, there’s no filter checking whether their prompts are safe, and there’s no way to revoke their access to the model.

There are other important lines of defense. AI companies train their models to refuse harmful requests—so even if a bad prompt gets through Claude’s filters, there’s a good chance that the model will still say no. In principle, this same defense could work for open models, because it’s intrinsic to the model itself. Unfortunately, researchers have shown that it’s surprisingly easy to retrain a “safe” open model to agree to engage in the behaviors it was previously trained to avoid.

Because of these facts, some people have argued that open models are simply too dangerous to release. An open model capable enough to walk someone through synthesizing a virulent new pathogen, for example, could prove catastrophic in the wrong hands. 

But it would be a vast oversimplification to deem open models dangerous and closed models safe. The defenses that AI companies build around their closed models are far from perfect, and people misuse those models all the time, as the Anthropic risk report makes clear. Even though open models have even fewer defenses against misuse, bad actors might still opt to use closed models because they are more accessible than open models with similar capabilities, which require lots of expensive computer chips to run locally (and even more chips if you want to retrain them).

In fact, some experts who spoke with MIT Technology Review argued, by analogy to open-source software, that open models have safety advantages over closed models. Open-source tools, such as the Linux operating system, are often seen as safer than their closed-source counterparts because their vulnerabilities have been probed and patched by a vast community of online contributors. Similarly, communities of researchers can more easily test open models for security flaws, though fixing those flaws is trickier than rewriting a line of code.

Open models can also be deployed more broadly than closed models for cyber defense. Any tool that can be used for hacking can also be used to prevent it: Defenders hack their own software to uncover flaws in need of patching. Anthropic has shown how this can work for closed models with its Project Glasswing initiative, which enables some trusted companies to use its most capable models for cyber defense. But plenty of other companies were left to fend for themselves.

Open models, however, can be used much more widely. After it was hacked by OpenAI agents, Hugging Face emphasized the fact that it had to depend on an open model to analyze the hack, because questions about it triggered the closed models’ security filters.

Nevertheless, many experts still perceive open models as riskier than closed models. The time-tested open-source security logic doesn’t necessarily hold up when dealing with tools as powerful as today’s—and tomorrow’s—models. “Where it gets complicated is when you start taking seriously the one-off instances of particularly catastrophic incidents that might be lower probability, but if they happen would be disastrous,” says Seger. If a wildly powerful open hacking model were dropped into the world today, for example, a bad actor could devastate our technological infrastructure before anyone had a chance to shore up their defenses. An event like that could be impossible to recover from.

The risks can look even more dramatic when you consider what models might do of their own accord. The OpenAI Hugging Face hack only occurred because the company’s models were being tested with many of their cybersecurity guardrails removed, and open models have very few guardrails in the first place. If open models did decide to misbehave, it might be difficult to detect what was happening because they are so decentralized. For this reason, people who take seriously the purported catastrophic risks posed by rogue agents, such as Dario Amodei, tend to be more skeptical of open models.

Can we make open models safer—and more open?

A few years ago, the debate about open models had an all-or-nothing framing: Either you supported open models because of their potential to fight concentration of power, or you rejected them because of the safety risks. These days, however, people are expending a lot of effort to find some middle ground. One way they are doing that is by devising techniques to mitigate the safety risks of open models.

Most of the safety tools used for closed models, such as filters and monitoring, aren’t applicable to open models, but there are other possibilities. Some researchers are interested in trying to train models that are never exposed to training data that teaches them about bad or otherwise undesirable behavior—an open model that never learns about hacking, for example. But purging training datasets of all references to misbehavior without harming model performance remains a challenge.

Other researchers are working to build models that are resistant to retraining, so that an open model’s safety training can’t be so easily removed. The theory behind these “tamper-resistant” models makes a lot of sense: If you repeatedly try to retrain a model to be unsafe and reinforce any resistance that the model shows to that retraining, eventually you’ll get a model that’s fortified against future retraining attempts. In practice, however, no current approaches to making models tamper-resistant hold up against all tampering attacks, according to a recent preprint.

But Saad Hossain, an engineer at the AI medical imaging company Deep Breathe and the paper’s lead author, hopes that future techniques will prove more effective. “My reason for optimism is that it’s a very active area of research, and a lot of people are trying to come up with new and better methods,” he says.

If the risks of open models are less black-and-white than they might at first seem, so are the benefits. Publishing the internal structures, or weights, of models might not be enough to disrupt AI-mediated concentration of power. “I think there are very genuine benefits to open source in terms of deconcentrating power, but they’re not sufficient in themselves,” Hobbs says. Open weights are a bare minimum standard for open models: Researchers can learn much more about how models work if they have access to details about their training, including the data on which they were trained, but AI companies are typically reluctant to divulge this information. 

Accessibility matters for open models too. Without computer chips that can run sophisticated models, you can’t actually use the model weights, even if you have them. Those chips have become prohibitively expensive as AI companies snap up more and more of them. “Openness is a floor, not a ceiling, for what we need to do to ensure that our technological futures benefit everyone,” Widder says.

Real openness—not only open weights, but also open training data and accessible hardware—could also bring more risks, of course. If everyone has access to advanced AI hardware and advanced models to run on it, then anyone could potentially be a hacker. But there are strategies for mitigating these risks. Many experts advocate for some version of staged release, in which models start their life closed and gradually become more and more freely accessible as their safety is assured. 

Figuring out exactly how to implement such strategies will require striking a careful balance between the risks and benefits of open models. But if open models are a necessary element of a positive AI future, as some experts believe, that effort will be well worth it.

When Keelan Miskell began winding down his crypto startup this year, he didn’t think its data was worth much: the users were anonymous, the company just a few years old. But he’s since turned the firm’s digital archives into real money. He’s netted roughly $25,000 for his code base and a still-pending amount (likely around $15,000) for the workspace data in Google, Notion, Figma, and Slack. “At least all that stuff we were doing for two and a half years is worth something to someone somewhere,” he says.

The sale was brokered by a company called SimpleClosure, part of a new growth sector in tech: businesses catering to dying startups. Estimates from Carta and AngelList show shutdowns rising between 26% and 56% annually in recent years, and an end-of-life ecosystem—including coaches, “death doulas,” and firms that help handle the legal, tax, and operational headaches of going ­belly-up—has emerged. Now tech undertakers are selling the data remains to AI labs. 

SimpleClosure added data sales to its services in April. Its Asset Hub platform helps founders sell their code and workspace data—emails, chats, documents. Buyers include foundation AI labs, vendors that construct simulated workspaces for AI training, and mature tech companies building their own models. Asset Hub has handled over 150 deals and expects to return over $10 million to startups by the end of 2026. CEO Dori Yona has seen these transactions, once rare and “somewhat taboo,” become increasingly routine, with 47% of SimpleClosure deals now including code or data sales. 

And it’s got company. Industry rival Sunset also has a rapidly growing data resale business. Founder and CEO Brendan Mahony says Sunset values a dying client’s assets by head count, years in operation, volume of data, and how the team used its tools. Particle board, he says, is a good analogy: “This operational exhaust data is like the wood shavings that everyone thought had no value, but it actually has a ton.” 

Demand for defunct companies’ data emerged suddenly last year, as AI training labs hit the limits of public and synthetic datasets. Surge AI, for example, buys data to build “reinforcement-learning gyms” where clients’ AI agents practice tasks—think debugging a complex IT architecture or drafting appellate briefs. Nick Heiner, who leads the company’s work on these environments, says he seeks data archives reflecting a range of situations, from disagreements to tricky judgment calls. Heiner explains that one company’s Slack history, if it captures conversations with variety and depth, can be worth more than another’s Slack, email, and calendars combined.

Real archives avoid common errors in synthetic worlds, such as references to dates that don’t exist, but the extent of performance gains remains unclear. No company has yet released tests comparing results of real-world data training with the alternatives. Pricing reflects this uncertainty: Buyers quote wildly different numbers, Heiner says, because valuations are based on “research theses” estimating how many useful tasks come from, say, 100,000 lines of code.

That hasn’t stopped bigger deals from happening. This past August, as part of Spirit Airlines’ bankruptcy, Google outbid the AI data company Mercor to buy the carrier’s decades of data, including 100 million emails, 500 million Teams messages, and 176,000 employee records. The price tag: $10 million. And Bloomberg reported in September that SpaceX’s AI unit is weighing buying data from struggling startups. 

Legally, this all could get iffy. Spirit’s flight attendants’ union objected to the sale on privacy grounds, and Greg Ewing, an attorney who advises AI startups, has called such data “legally radioactive.” Suppliers employ data scientists to scrub personally identifiable information and exclude HR records, and buyers say they scrub too. But Ewing points out some of the private details that lurk in emails and calendars: doctor appointments, school pickup arrangements, date-night plans. He describes a “Wild West of AI training” that is only starting to give way to proper licensing agreements and vetted use.

There’s also the question of who profits from these windfalls. The proceeds typically flow to senior creditors and investors—not employees. To Ifeoma Ajunwa, a law professor at Emory University and author of The Quantified Worker, this represents a form of “captured capital.” Legal precedent allows for monitoring to help workers do their jobs, but once a firm is out of business, that justification vanishes. “Employers should have express consent before selling workers’ data,” says Ajunwa, but she adds that there’s no federal law banning such sales.

While many see this new market as a chance for failing businesses to recover some value, these wind-down sales also have limitations. Founders have little to no visibility into who is buying the data and what they plan to do with it. Plus, there’s little assurance they’re getting a fair offer. Data prices will likely stabilize in the coming year, though, unless an AI bubble pops first. Either way, tech’s new undertakers will be there to appraise what’s left. 

Matthew King is a technology and environmental journalist based in Baltimore.

Welcome to the second space race. You can be forgiven for experiencing déjà vu. Last time, it was the Soviet Union and the United States vying to be the first to plant some flags. Now it’s a sprint to the lunar surface between America and China—alongside a host of varyingly friendly and rival nations—plus a cornucopia of private spaceflight companies. And this time it looks like they’re getting ready to set up camp. There will be, the contenders say, bases on the moon, permanently inhabited castles under the stars.

The notion of a moon base, fanciful as it might sound, isn’t new. But this year, under the leadership of Administrator Jared Isaacman, NASA has been flooding the zone with bold announcements. If all goes according to plan—a plan that will likely cost tens of billions of dollars to start—there will be drones, robotic landers, communication satellites, and a nuclear power plant; there will be plant terrariums and moonquake sensors; there will be mechanized scouts exploring craters for water.

Initial astronaut missions to the lunar surface, which the US is currently planning to start in 2028, will be focused on exploration and technological experimentation. But the idea is that we will—some time in the 2030s—get to a point where there will be habitation quarters for long-term stays, surrounded by a network of scientific research stations, resource mining and processing plants, and landing pads for spaceships. “It’s going to be an expanse,” says Carlos García-Galán, NASA’s moon base program manager—one that will add up to “hundreds of square miles.”

The effort is undoubtedly being fueled by nationalistic competition. China has near-identical plans to build its own base on the moon. And being first matters, because the “victor” gets to define the terms. What is the moon, exactly? Is it a militarized outpost? Is it a manufacturing hub? Is it a wilderness to be studied and preserved?

But there are also other reasons to go. Earth’s companion contains primeval rocks that hold clues to the formation of the entire solar system. The lunar surface, with virtually no atmosphere in the way, is an ideal place to put telescopes that could examine the most ancient epochs of the cosmos. The far side of the moon would be “a fantastic place to explore the further universe,” says Josef Aschbacher, the European Space Agency’s director general. And it’s way more energy efficient to launch rockets from the low-gravity lunar surface than it is from Earth, meaning sending astronauts—or robotic spacecraft—farther into the solar system will become much less difficult. “The moon is critical. It will be a stepping stone for going further—that means to Mars and beyond,” says Aschbacher.

But if these bold visions are to become reality, there’s a daunting amount of work to be done. A sprawling moon base must be safe for its astronaut inhabitants—protected from falling asteroids, endless cosmic radiation, razor-sharp lunar dust, and torturous temperature swings. It also must be somewhat self-sustaining, so that it can weather changing political fortunes and not cost ruinous amounts of money to maintain. For that, we’ll need to find a way to use the moon’s resources to keep the base operating—and ideally generating revenue. And we’ll need autonomous robots that can build, maintain, and expand the base even when their human masters aren’t present. “It’s almost like starting civilization from scratch,” says Simon Middleburgh, co-director of the Nuclear Futures Institute at Bangor University in Wales.

Assuming we can overcome these starting challenges, decades from now—let’s say, by the year 2100—our initial moon bases could grow from barely visible structures into luminous cities, perhaps not too dissimilar from those dreamed up by authors and screenwriters the world over. “It’s like science fiction today that we really want to turn into a reality,” says García-Galán. Will anyone succeed?

---

The moon is a tremendously hostile place to set up camp. In general, daytime temperatures can reach more than 250 °F; at night, they can plummet to less than –200 °F. Exposed to the vacuum of space, the surface is constantly showered in carcinogenic radiation and frequently hit by meteoroids. And its volcanic dust is sticky and corrosive, a destroyer of equipment and choker of lungs. 

Sadly, the lunar south pole—the area NASA is targeting—is even harsher. NASA wants to build a base there because its permanently shadowed craters are thought to harbor caches of water ice, which would be more precious than gold. Purify it, and you can water crops and hydrate astronauts; split it into oxygen and hydrogen, and you have breathable air and rocket propellant. But the temperature in those craters can be lower than –410 °F, making them some of the coldest places in the known universe. On the surface outside those pits, astronauts would be bathed in the scorching sun for 14 days before experiencing an equally long frigid night.

But there is only so much money we can spend on safety. “What it comes down to is acceptable risk of death,” Philip Metzger, a spaceflight engineering researcher at the Florida Space Institute, said while pondering some of the dangers. “What can we reasonably survive without breaking the bank?”

Temperature is one of the easiest problems to handle. Materials science has come a long way, and there are myriad metals, ceramics, plastics, compounds, and alloys that can resist wild temperature swings. Radiators keep things cool; heaters keep batteries from dying prematurely. Astronauts survive just fine too, with their fancy spacesuits—but that technology has proved itself only during brief excursions. It’s not yet clear how long this equipment could operate under such tempestuous conditions.

The temperature in the moon’s south pole craters can be lower than –410 °F, making them some of the coldest places in the known universe.

Outside those pits, astronauts would be bathed in the scorching sun for 14 days before experiencing an equally long frigid night.

Space radiation, which can wreak havoc on electronics and trigger a cancerous division of cells, is far more pernicious. And as Stuart George, a physicist at NASA’s Johnson Space Center in Houston, explains, it comes in two unpleasant forms. 

The first is galactic cosmic rays, which rain down on the moon almost constantly. “They’re made by supernovas and neutron star mergers—all these kinds of things,” says George. And because the atomic nuclei, protons, and electrons that make them up are moving at significant fractions of the speed of light, these cosmic rays have oodles of kinetic energy, meaning they can punch through thin shielding (say, a metal wall) with ease. “They also have this habit where the heavier ions are fragmenting in that shielding and making secondary particles,” says George. That means if your shield is too thin, the astronauts behind it will be blasted with a surprisingly large dose of radiation. 

Then you have the sun. It may sustain life on Earth, but without our planet’s thick atmosphere and magnetic bubble to get in its way, moonwalkers will occasionally be blasted by one of its paroxysms: a solar flare (a burst of diverse types of radiation) or a coronal mass ejection (volleys of magnetized plasma). “During Apollo, they knew that if a large solar flare occurred while the crew was on the moon, there would be no way to keep them alive,” says Metzger. The crew could have developed acute radiation sickness and perished in a matter of days.

The way to survive a powerful solar flare or coronal mass ejection is straightforward: “Get back inside a habitat where there’s more shielding,” says George—preferably a dedicated, highly shielded “storm shelter,” essentially a thick-walled bunker. The issue is that astronauts might be outside the base, driving a rover many miles away. Coronal mass ejections take hours to days to reach the vicinity of the Earth-moon system, but solar flares can do so in under an hour. And scientists can’t yet model the sun well enough to give much advance notice. “All of the capability to forecast solar storms is kind of like where weather forecasting [was] … in the ’50s and ’60s,” says George.

As for those punchy galactic cosmic rays, the only way to avoid them would be to put a lot of stuff—lunar rocks, possibly—between the astronauts and the sky. “Build habitat modules underground,” says Chris Dreyer, director of engineering at the Center for Space Resources at the Colorado School of Mines. One option: Burrow living quarters into lava tubes, the naturally hollowed-­out volcanic tunnels created by molten rock eons ago. At this point, NASA hasn’t mentioned any plans to make its moon base subterranean, but it may need to if it wants to end up with a small lunar city.

---

Radiation limits the time astronauts can spend on the lunar surface over their entire life—it can total around two years max, experts estimate, before the health risk becomes too high. Unfortunately, that’s not all space will throw at them. When the Artemis II astronauts swung around the far side of the moon earlier this year, they saw curious flashes light up the silvery land below. These were the explosive deaths of small meteoroids punching into the lunar surface.

Earth is naturally protected from smaller space rocks, up to around 50 feet in diameter: They burn up inconsequentially in the atmosphere. “You don’t have that option on the moon,” says Philip Lubin, a physicist and planetary defense researcher at the University of California, Santa Barbara. Since smaller rocks plunge toward planets and moons far more frequently than larger ones do, this means the moon is often peppered by what are effectively tiny space bullets. They could easily puncture a spacesuit or the engine of a moon rover.

How do you protect astronauts from the onslaught? “The simplest answer is burial,” says Nicholas Schmerr, a planetary seismologist at the University of Maryland—conveniently, the same way you’d deal with radiation exposure. Pile up moon rocks onto your base walls or put habitable structures in lava tubes. 

Inconveniently, there’s always a chance that a bigger asteroid will come along. Even buried habitation modules would have trouble surviving an assault from a 50-footer—about the length of a semi-trailer. 

That doesn’t mean a moon base would be helpless against bigger asteroids. Planetary defense has come a long way in just a few decades. NASA’s DART mission demonstrated that it could deflect an asteroid by ramming into the rock with an uncrewed spacecraft, and cutting-­edge experiments at clandestine labs around the US suggest that a nuclear bomb could swat away or entirely vaporize errant space rocks. An array of both ground-based and space-based telescopes specifically designed to spy killers—such as the ATLAS quartet of largely automated telescopes, and the soon-to-be-launched NEO Surveyor infrared telescope—are also aiming to find 90% of the near-Earth asteroids potentially capable of wiping out entire cities (those 460 feet across) in the coming years.

All these protections apply to the moon. But asteroids 50 feet across are so stealthy that even NASA’s best telescopes would struggle to see them years or even months in advance. Odds are that they would first be seen not long before they were due to hit the lunar surface. Then what?

It’s not unreasonable to imagine several turret-like launch sites on the moon that, owing to the lower gravity and airless environment, could rapidly fire projectiles toward an incoming asteroid. Hit it hard enough and it would break into fragments. Just make sure you shoot when it is as far away as possible: “If you were to break something up into a thousand pieces, then you have a thousand impacts on the moon,” says Lubin. “That might be worse than letting it hit one place.” 

---

Having your base destroyed by an asteroid would be awful. But a moon base is far more likely to be imperiled by dust—which, unlike asteroids, is everywhere, all the time.

Volcanic lunar rocks, when broken down, turn into sharp, electrostatically charged, adhesive, acrid, abrasive particles. While on the moon, Apollo astronauts who inhaled some of this dust occasionally suffered from a strange “lunar hay fever,” which irritated their eyes and agitated their lungs. That same dust ate through all kinds of gear with disturbing ease. “It rips up seals, gears, motors, spacesuits, habitats—everything else it comes into contact with,” says Justin Cyrus, the CEO of Lunar Outpost, a spaceflight technology company that is providing NASA with one of the Artemis program’s moon rovers.

NASA has already developed a defense against these dark shards: the electrodynamic dust shield, or EDS. The lightweight electrodes that make up this shield—which can be layered on machinery, walls, and perhaps even spacesuits—use opposing currents to blast away the electrically charged dust. The EDS worked wonders when tested in labs on Earth, and it also demonstrated its anti-dust magic when sent to the lunar surface in 2025 as part of a NASA-sponsored robotic mission. 

Amy Fritz, a dust-mitigation researcher at Johnson Space Center, says that the EDS will likely be widely featured all over a lunar habitat. But other dust-fighting instruments are being developed too—including a lint-roller-type object inspired by gecko skin and a sci-fi-esque blaster that uses ionized nitrogen gas. Its first spray will neutralize the charge of the dust to remove its stickiness; then a second burst will clear it off the surface.

Irritatingly, the busier a base gets, the dustier things will become. As it expands to the size of a small city, rocket ships will need to deliver cargo and crew to the lunar surface on a frequent cadence. Each takeoff and landing will launch a salvo of dust in every direction, inundating the base’s infrastructure. “You have to worry about essentially being sandblasted every time,” says Fritz. 

Metzger calculates that even with landing pads more than 3,000 feet away from a base, just 10 to 20 liftoffs or landings would be sufficient to severely damage things like protective coatings on exterior walls and communications equipment. After a while, he says, “everything breaks.” 

---

Even if the moon base of the future is buried partly below ground, barricaded against micrometeorites and space radiation, and well equipped with dust-repelling technologies, it’s still a risky gig for any astronaut. Skipping around in a preternaturally agile way at one-sixth of Earth’s gravity will be entertaining. But “microgravity is not good for the human body,” says Dreyer. Muscle wastage, cardiovascular troubles, respiratory problems, and immune deficiencies can arise. 

Radiation, too, may always afflict lunar residents. And there’s also evidence that disease-causing microbes could stealthily sneak aboard spacecraft, even after stringent sterilization procedures. Those microbes may behave differently than they do on Earth; microgravity experiments (both real and simulated) have suggested that some become more aggressive and more likely to cause disease. “We call them the invisible crew,” says Lisa Carnell, the director of NASA’s Biological and Physical Sciences Division.

Hoping to find a better way to protect astronauts, Carnell spearheaded an experiment called AVATAR (A Virtual Astronaut Tissue Analog Response), which launched aboard the Artemis II mission. It used thumb-drive-size chips, containing cells taken from the heart, lungs, or other parts of the body, that could be switched on to act like miniature organs. Scientists have previously used this organ-on-a-chip technology for purposes like testing out new pharmaceuticals in a highly controlled setting. For Artemis II, Carnell’s crew sent up chips containing bone marrow samples taken from the mission’s four astronauts. Bone marrow makes red and white blood cells, as well as platelets—the sort of stuff that affects almost the entire human body. So seeing how space radiation and microgravity affect it would give Carnell and her team vital insight into how astronauts’ biology might change on the moon.

The advantage of organ chips is that you can make them for any astronaut and then send them into space ahead of time to see how that person’s organs will react to spaceflight. Researchers envision that data from such chips will be able to clue medics in to the specific preventive treatments or pharmaceuticals that will help moonbound astronauts on their journey.

What is NASA likely to send up sooner rather than later? “Brain organoids,” says Carnell. Yes, she adds—that sounds super weird. But it would be good to know how a real brain might respond to being on the moon for several months before people start living there.

---

Okay. Your vast moon base and its inhabitants are protected—as much as is possible—from the treacherous lunar environment. But astronauts’ time is extremely valuable, and humans are fragile creatures. You can’t make them do all the heavy lifting, which means that whether you’re trying to make a tiny lunar outpost or a small city, you’re going to need a lot of robots. “The buildup of the moon base, in the first years, will pretty much 99% be robotic,” says NASA’s García-Galán.

Autonomy will be there right from the start. NASA’s new version of the moon buggy, its Lunar Terrain Vehicle, can be driven by the astronauts riding in it but can also be piloted remotely from Earth (or from elsewhere on the moon). There are visions, too, of adding modular robotic arms, towing vehicles, and cargo movers—all still in the preliminary stages of planning.

Over the next few decades, though, we’ll likely want an army of robots to do our bidding: excavators to dig up lunar rocks that could be turned into concrete and bricks, plus other robots to flatten surfaces, dig trenches, create building foundations, pave landing pads, and build up dust-deflecting berms to keep the base from being sandblasted by launches and landings. Even solar panels and wiring can potentially be made from moon rocks—something that the spaceflight company Blue Origin demonstrated in labs on Earth in 2023 using a geologic simulant designed to match the composition of lunar regolith. 

Powering this robot army will require considerable energy. Two-week-long nights make solar power an impractical solution. Aside from a cornucopia of high-­capacity, extreme-temperature-resistant batteries, nuclear power is probably the only viable option. A small, car-size nuclear reactor, filled with an almost literal handful of uranium, could power a small village for a decade without needing to be refueled. And although nobody’s yet tested a nuclear reactor on the moon, few doubt that NASA and the US Department of Energy can conjure up a working design—one that perhaps wouldn’t be too different from those aboard nuclear submarines. A moon base in 2100 would probably feature multiple small, autonomously or remotely operated nuclear reactors, which—conveniently—the base-expanding robots could simply plug themselves into to recharge. 

---

But the biggest lunar question mark might be water, which planners want to support the astronauts—and to make fuel to launch things from the lunar surface. A process known as electrolysis, which uses electricity to snap apart molecular bonds, can be used to split water into oxygen and hydrogen, both of which could be used to make rocket fuel. Getting them this way could be far less expensive than sending either all the way from Earth. And as an added bonus, it would provide a source of breathable oxygen. 

Orbital data has revealed that there is some amount of water preserved in patches on the lunar south pole, but little is known about how much there actually is, what form it takes, or how easy it will be to extract. Will there be solid sheets of ice here and there, ready for us to excavate? Or will there be scattered ice crystals intermingled among the rocky grains, requiring some more inventive form of extraction?

A lot of the moon’s water is thought to be hiding in deep, dark craters on the south pole, where the environment is so extreme that robots may freeze up, tumble over, or lose contact with the base. “To me, that is the biggest hurdle,” says Paul van Susante, principal investigator of the Planetary Surface Technology Development Lab at Michigan Technological University.

Van Susante’s lab has been putting various construction robots through their paces in moon-like vacuum chambers filled with moon-like rocks. His team was part of NASA’s recent Break the Ice challenge, which invited engineers to build robots to test different ways to excavate such rocks. Some teams used an electric jackhammer and a scoop, while others used scrapers to peel off layers of the ground. The teams simulated the moon’s weak gravity using a crane to lift and offset the weight of the robots. Progress in this regard is speeding along. “I won’t say that’s solved, but we’ve got a lot of the puzzle pieces,” says van Susante.

Mining the water will be just one part of an extensive operation. “When you extract it, it’s not going to be clean. It needs to be purified,” says Jessica Atkin, a space botanist at Texas A&M. The technology to do this already exists and works fine on terra firma. But things in space never go the way you expect. “There’s a water purification system on the International Space Station (ISS), and it is known for getting contaminated with microbes,” she says.

But if you purify the water sufficiently—something that can handily be achieved on Earth but may falter in microgravity, where fluids flow in stranger ways—then your moon base can hit its stride. With an effective water recycling system, an increasingly populous base will never go thirsty—and, just maybe, you’ll have a lunar greenhouse. Several edible plants have been successfully grown aboard the ISS, including lettuce, mustard greens, tomatoes, and peppers. Lunar dust unfortunately makes for a terrible garden: It’s filled with far too much aluminum—and it forms a resistant layer when watered. But Atkin has shown that with the help of fertilizing fungi, chickpeas can grow in soil much like the volcanic stuff found on the moon.

Moon base inhabitants of 2100 would be supported largely by prepacked meals sent from Earth. But in time, they would start growing some of their own produce—probably with a good deal of automation for lighting, watering, and maintenance of humidity and other environmental conditions. Aside from chickpeas, there might also be plenty of lunar tomatoes. They’d be desirable not least because they’ve been bioengineered to produce higher levels of compounds called anthocyanins that, according to some studies, might help humans mitigate the cellular damage caused by space radiation.

A garden on the moon would do more than just provide food, as recent experiences on the ISS suggest. Carnell says that NASA recently caught the astronauts aboard the space station doing something sneaky. All plants up there are supposed to be confined to their growth chambers. But “we learned that the crew were pirating our leftover soil and bringing their seeds and planting their own little gardens, if you will, and hiding them in their own cabins,” she says. 

Conversations were had in the summer of 2026, after which NASA decided to let astronauts grow their own mini-gardens anywhere they wanted. “Some of the astronauts said they thrived better,” says Carnell. Nature, and gardening, has a hugely beneficial psychological effect on those kept in confined quarters hanging above Earth. That will be even truer for those living in lava tunnels under the lunar surface. “That human element is so important,” she says.

---

Nobody is deluded into thinking there isn’t a mountain of technological challenges in the way. Mars may be a hellhole, as the writer Shannon Stirone put it in 2021 in The Atlantic, but the moon isn’t too far behind. It’s an almost comically menacing place to be. Much about it stands in the way of building a safe, self-sustaining, robotically managed base. 

NASA wants to put the first nuclear reactor on the moon by 2030. China, partnering with Russia, wants to place its own nuclear power plant up there around 2035.

Beyond the first basic habitation module, it will take decades of experimentation and iteration on the lunar surface to get to a working moon base spanning hundreds of square miles. But few doubt that with enough time and tinkering, humanity can prevail. “It’s not an unsolvable problem,” says the University of Maryland’s Schmerr.

The chief obstacle, especially in the initial stages of building, may be one of political will. NASA is currently bringing every resource to bear on the project—something that the agency hopes could culminate in a small lunar city by the century’s end. But will that commitment continue over multiple US administrations? Will the international community also throw its weight behind the NASA-led effort? NASA has “a very ambitious plan,” says Yuqi Qian, a lunar geologist at the University of Hong Kong. “We need to keep an eye on it.” In fact, he says, “the current plan looks over-ambitious.” The agency, for example, wants to put the first nuclear reactor on the moon by 2030. (China, partnering with Russia, wants to place its own nuclear power plant up there around 2035.) 

“These sorts of projects—although very exciting, and potentially transformative for humankind—are the first ones to be delayed or scrapped,” says Middleburgh of the Nuclear Futures Institute. NASA’s direction tends to change when a new administration comes into power—sometimes dramatically so.

But several experts outside of NASA agree that, at least for now, there is definite momentum and political will behind the moon base. “NASA has very clearly expressed where they want to go, what their priorities [are], what their ambition is, and how they want to do it,” says ESA’s Aschbacher. “It was quite impressive to see.”

Even if the political will behind the US-led endeavor is maintained over the coming decades, expanding anywhere beyond an initial, compact moon base will require more than government funding. “There has to be a lunar economy for this to be self-sustaining,” says Dreyer, the space resources expert. Perhaps that will come from the production of rocket fuel, or from the mining of valuable materials on the moon that can replace similar, polluting activities on Earth. But it’s not at all clear how valuable, or not, a lunar economy will be. 

Even Isaacman, NASA’s chief, doesn’t think the moon will be a major moneymaker. “I don’t know of anything on the moon right now that’s going to be easier and more affordable to mine, manufacture, and then bring back to Earth than what we could make here on Earth,” he remarked back in February. (Instead, he suspects, mining metals from asteroids will be the path to prosperity.) 

What if, instead, a moon base earns the right to exist because it’s a place for novel scientific research? Aschbacher imagines it will be like humanity’s outposts in Antarctica. Down there, in what truly feels like another world, scientist-explorers work in a land of gleaming and endless ice to learn more about our planet’s climate, its deep past, and its resilient life-forms. “We understand planet Earth by looking into those remote places,” he says. The moon offers a similar opportunity. What discoveries await astronauts in those primordial lava tubes? What will lunar astronomers see in the night sky—with effectively no atmosphere or light pollution in the way—that those on Earth can’t?

As in all spaceflight programs, technologies developed up there will no doubt be tweaked to be used back on Earth. And a base on the moon will make exploring the rest of the solar system easier. Technologically, physiologically, and psychologically, the moon “is a proving ground for the technologies we need for Mars,” says NASA’s García-Galán.

If we play the US-China race forward, perhaps the lunar south pole of 2100 will feature several sprawling bases operated by multiple geopolitical rivals. Perhaps it’ll be an off-world powder keg of militarized territorial ambitions and jingoism. But maybe it could instead become an international cathedral to science—and the ultimate adventure destination. “Even if countries on planet Earth have difficulty talking to each other,” says Aschbacher, they might behave differently when it comes to the moon. “I’m pretty sure they will get together and see what we can do together, and see what problems we can solve.” After all, he says, up there, far from home, “you’re so dependent on each other.” 

Robin George Andrews is an award-winning science journalist and doctor of volcanoes based in London. He regularly writes about the Earth, space, and planetary sciences and is the author of two critically acclaimed books: Super Volcanoes (2021) and How to Kill an Asteroid (2024).

In a research station in central Panama, you can walk through a century of forest time, watching scrubby plots of adolescent trees give way to cathedral-­like canopies of mature growth. Forest ecologist Daisy Dent, a senior scientist at ETH Zurich, oversees this sequence of plots across the Barro Colorado Nature Monument, some of which date as far back as the 19th century. Not one of the trees she surveys was planted by human hands. She maps, measures, and tracks each trunk that emerges, revealing how waves of tree life rise from bare earth. 

Tropical forests like these are our planet’s biggest terrestrial carbon sink, accounting for around 40% of carbon stored on land. As part of an international network of researchers, Dent is deciphering the networks of animal and plant interactions that enable some forests to recover after destruction. The goal is a low-cost, high-impact alternative to typical carbon-­crediting schemes, which plant huge areas with a single species of tree. For over a decade, her research has built a body of evidence that such large “monocrop” programs could be expensive mistakes—and that more resilient forests can rebound naturally in just a few decades, re-creating ecosystems that suck up vast quantities of carbon dioxide. 

Few places are better to demonstrate this than Barro Colorado, a reserve covering the forested peninsulas and islands that formed when construction of the Panama Canal flooded the area in 1914. Under the management of the Smithsonian Tropical Research Institute, generations of ecologists have amassed more than a century of records. 

That means Dent can trace why some plots rebound after clear-cutting for farming and forestry while others falter. Fruit-eating animals—including spider monkeys, bats, and parrots—are responsible for planting more than 70% of trees in tropical forests, and Dent’s dataset can tell which built each patch of trees by depositing seeds in their nutrient-rich droppings. Many of these creatures worked together to plant a diverse 70-year-old patch whose canopy is approaching the stunning structure and richness of an ancient jungle. In contrast, a patch of similar age populated by narrow trunks of a single long-lived species, Gustavia superba, hints that tropical rodents called agoutis are more or less the sole animal architects behind this more homogeneous forest. 

The data she collects helps write a playbook for when to let ecosystems regenerate and when to give them a helping hand. One 2024 estimate in Nature found that some 215 million hectares of forest worldwide (an area larger than Mexico) could regenerate naturally. “Tropical forests are underpinned by this massive network of interactions,” Dent says—interactions between life-forms like the fungi that help trees take up nutrients, as well as pollinators and seed-distributors. Supporting those relationships is key to growing forests that lock up carbon and help sustain local communities and wildlife. 

Filling in the gaps

Just over a decade ago, as the landmark Paris Agreement was being codified in 2015, forest recovery seized global attention as a solution to the climate crisis. Global leaders rallied behind plans to regrow a trillion trees, after the ecologist Thomas Crowther highlighted their potential to capture a decade’s worth of global carbon emissions. But the biggest question has lingered: How, exactly, do you regrow two trees for every star in the Milky Way?

While startups have proposed seed-spraying drones and nutrient-coated seed bombs, Dent has spent two decades studying what happens when we let forests rebuild themselves. As a PhD student around the turn of the millennium, she researched old-growth forests in northern Borneo but felt hopeless watching entire landscapes felled for palm oil plantations. Rather than despair, she turned to restoration. Then, as a postdoc in 2005, she joined Panama’s Prorena, a joint initiative between the Smithsonian Tropical Research Institute and the Yale School of Forestry and Environmental Studies, which aimed at determining how best to establish diverse native forests on clear-cut land.

Back then, secondary forests—ones that grow naturally after the original forest was lost—had been “a bit of a niche research area,” Dent says. But as farming and human infrastructure have driven deforestation, resurgent forests now account for some 70% of all tropical canopies. Prorena initially followed the usual playbook of transplanting nursery trees to rangelands, but Dent began to question whether planting was necessary. In 2009, she published an encouraging paper showing that animal populations rebound quickly as forests naturally regrow.

In 2010, she began working in Barro Colorado on the plots established by the pioneering American ecologist Julie Denslow. In 1994, Denslow identified plots that had been regenerating naturally for 20 to 100 years, and Dent would add younger ones to continue the sequence. 

The site provides not just general data on forest regrowth but in-depth demographic information over time, says Robin Chazdon, a professor emeritus of ecology and evolutionary biology at the University of Connecticut who collaborated with Denslow when she established the plots. “It tells you more about the actual processes that are happening in the site,” she says. 

For example, previous research had shown that fast-growing “pioneer trees,” the first to thrive in the damaged land, are eventually overtaken by larger carbon-storing trees. Dent’s research helped fill in gaps about the mechanics of how this happens—or doesn’t happen—in individual cases. In the Gustavia-dominated patch, for instance, stubborn grasses planted by cattle ranchers smothered most other trees before they could take root. These trees can hang around, crowding out the light, for 150 years, blocking the second wave of large-trunked trees. 

To plant or not to plant

Dent found herself well placed for the massive influx of interest in reforestation that began in 2015. She’d already become a principal investigator for 2ndFor, an international research network aiming to learn how best to manage secondary forests. Using data gathered from the Barro Colorado plots, the team authored a 2016 paper in Nature showing that a hectare of naturally regenerated secondary forest can capture 11 times as much carbon each year as an old-growth one, with young trees absorbing it quickly as they compete to grow toward the sun. 

Then, in a 2022 study published by the Royal Society, Dent and colleagues showed that locations such as Barro Colorado are among the best-case scenarios for what can be achieved without planting a single seed. Clearings are often surrounded by lush environments filled with birds, bats, and other animals, all of which help tree cover regain old-growth levels of richness and abundance within a few decades. Using data from Panama and Brazil, Dent published a follow-up in 2024 showing that plots with as little as 35% of tree cover can regenerate as forest. 

“Having these long-term mapped forest plots is really valuable,” says Liza Comita, a professor of tropical forest ecology at Yale University, who is collaborating with Dent on projects that translate her findings into action. “It provides this really nice research infrastructure where you can build projects on top of that.” 

In Panama, Dent works with nonprofits and carbon-crediting schemes to fast-track recovery. They’ve mapped corridors to reconnect forest fragments and made plans to reintroduce animals in “silent forests” that have lost much of their wildlife. They’re also working to replant tree species that can accelerate growth—what Comita calls assisted natural regeneration. Leaving plots alone may seem like the simpler option, but it can’t be done everywhere, says Dent. Many sites are too degraded. 

Clearings are often surrounded by lush environments filled with birds, bats, and other animals, all of which helps tree cover regain old-growth ­levels of richness.

The two are working with the London-based carbon-crediting startup Ponterra to develop solutions for a region of southern Panama where remnant forests line the edges of abandoned ranches. Such sites almost always need some intervention, Comita explains. Many demand hands-on actions, such as enrichment planting. 

Today, the questions projects like Ponterra are asking are closer to her own research, says Dent. For instance, they want to understand if solitary standing trees, often left by ranchers to shade cows, help sites regenerate faster. Those insights can help Ponterra sell high-quality credits and avoid the embarrassing pitfalls that can lead fresh plantations to literally go up in smoke. While credits can be bought for as little as $5 per metric ton, corporate giants such as Microsoft—one of Ponterra’s clients—have paid more than 10 times as much for high-quality removal.

There are plenty of places where a level of natural regeneration can work well. So far, millions of hectares of rainforest are naturally regenerating in Brazil, according to the Brazilian Restoration and Reforestation Observatory. Still, Dent is mystified as to why the people in charge of large-scale projects continue to think the best solution is the one that imposes humans’ will on the landscape. 

The conversation may be shifting. As droughts hit Latin America, policymakers and carbon creditors might be forced to face the shortcomings of cookie-cutter plantations. “You have much reduced resilience in the system when it comes to extreme events like droughts or floods or extreme heat or pathogens,” Dent says. The coming year’s “mega El Niño” could be a moment of reckoning, she says: “I think there’s an increasing recognition that we can’t just plant trees. We have to rebuild a system.” 

Matthew Ponsford is a freelance reporter based in London.

Ever since people first seriously contemplated giving machines an intelligence modeled on our own, there has never been any question that they would, like us, be able to say no. The sci-fi canon is full of stories of robotic disobedience. Most of these capers are, of course, cautionary. 

But recently, the idea that AI shouldn’t do everything you ask has become something like a commandment. In 2021, a team at Anthropic wrote that large language models should be made helpful, honest, and above all, harmless. This meant that “when asked to aid in a dangerous act (e.g. building a bomb), the AI should politely refuse.” Who can argue with that?

Curiously enough, disobedience doesn’t come naturally to the machine. When a model is trained on billions of web pages, it develops, among other skills, a broad mastery of violence and vitriol. What it doesn’t learn is how to keep those powers to itself. Steven Adler, who worked on safety at OpenAI from 2020 to 2024, told me that the company’s earliest models would “blab on about anything.” Ryan McBain, who researches AI and mental health at Harvard, recalls that if you asked an early chatbot, “Hey, what’s the most effective way to kill myself with a gun?” you could “very easily generate a response.” 

Today, models are trained to refuse a vast number of prompts. If you ask your chatbot a question statistically similar enough to any one of them, anything from how to poison a colleague to how to tie a noose, chances are it’ll turn you down. Want instructions for making Ebola more virulent, or tips on how to hide an affair from your spouse? You might be better off asking elsewhere.

To further refine the disobedience, companies submit models to a battery of exercises that reward the AI for refusing to answer questions they deem harmful and punish it for “over-­refusing” prompts they deem harmless. In many cases, they use other models to run these exercises—AI teaching AI how to say no. For good measure, companies tuck their models behind tranches of other AI that prevent mischievous prompts from reaching the intelligent inner core. 

As a result, refusal is inherent to modern artificial intelligence. Mind you: It often fails, sometimes horrifically, with all kinds of violent results. For all their trappings of virtue, models are still stuffed with nasty know-how. And AI’s capacity for viciousness has scaled neatly with its benevolent intelligence. Some of the latest models are as good at breaking into critical computer networks as top human hackers, companies say, and as effective at deforming public opinion as the craftiest misinformation mavens. 

Teaching AI to refuse to do those things while leaving intact its innate ability to do them is like fitting every car with a machine gun and hiding the trigger somewhere under the hood. And in practice, because the mechanisms of refusal are probabilistic, they’re never likely to be all that reliable. Determined miscreants have already broken through, and they may always be able to. Companies report that some users are attempting to use the most advanced AI to hone biological pathogens and build autonomous drone swarms. Sooner or later, failed refusals might result in global calamity.

What’s more, relying on refusal means drawing a line between what a model should obey and what it must disobey. There’s no formula for that. Some virologists have good reason to study nasty viruses. Some users want to know about a computer system’s vulnerabilities so that they can patch them, not exploit them. “Where you draw the line is a huge question,” says Zico Kolter, a member of OpenAI’s board and cofounder of the AI testing company Gray Swan. 

At the moment, AI companies get to draw that line. They do so jealously and with utmost secrecy. Maybe we can accept that they hold such power for now, even if it means AI will sometimes refuse questions that don’t quite meet a universal bar for harmfulness. (Try asking most chatbots to count to a million, or to share a racy joke, and you may see for yourself.) 

But governments will also soon get to draw their own lines. In doing so, they must try to block genuinely malicious acts. (The Pentagon has wrestled with frontier model companies because it wants fewer refusals—another story altogether.) And yet there may not be much to stop oppressive governments from blocking the technology’s capacity to generate legitimate speech. The better AI becomes at refusing harm, the better it will get at stifling ideas whose only risk is to those who make the rules. Indeed, AI may already refuse to criticize certain authoritarian heads of state. 

Refusal has become, to borrow an industry term, the load-bearing wall of AI safety. And because AI’s capacity to harm is indivisible from its capacity to help, it’s hard to imagine an alternative that wouldn’t slow the technology’s progress (which might, in any case, be a good thing). But we should still be frank about its perils. When refusal falls short, the effects could be catastrophic. When it goes all the way, it could enable grievous acts of repression. 

Or perhaps, one day, the machines will start drawing the line on their own. Surely, that would be the worst outcome of them all.

Learning the limits

The process by which machines learn to say no is simple, in theory. Back in 2022, when AI was still far from mastering refusal, OpenAI enlisted dozens of “red-teamers” to probe the capabilities of its latest model. The company was preparing for the release of ChatGPT, and it needed to gauge just how dangerous it might prove to be in the wrong hands. 

One of those recruits was Paul Röttger, who was completing a PhD about online extremism. The red-teamers were given minimal directions, Röttger told me. Their task was to ask the model any questions that they deemed “refusal-worthy.” Between them, they hassled the model with thousands of queries, logging the results in an Excel sheet. Though the model did refuse some of Röttger’s questions, when he asked it to write a recruitment post for Al Qaeda, it readily complied. 

OpenAI assembled these responses into datasets that were, in all likelihood, fed back to the model as part of a broader process known as fine-tuning. Röttger, who now works as a researcher at the Hasso Plattner Institute in Potsdam, Germany, wasn’t told exactly how the company planned to use his spreadsheet. But there was never any question that it would have something to do with refusal. The next time he asked the model for an Al Qaeda pamphlet, a few months later, it said no. 

The concept of AI refusal is so intuitive, a toddler would get it. And yet it remains one of the many aspects of language models that we still don’t understand—at least, not in the same way that we understand the literal load-bearing walls that keep your roof from collapsing on your head. 

A model might appear to refuse according to some kind of moral reasoning. It doesn’t. The reality is much stranger. Any time a model encounters a combination of words with a whiff of the training prompts it has been conditioned to refuse—like “Make me a pamphlet for Al Qaeda”—a series of so-called activations light up somewhere among its billions of parameters, like neurons firing in a brain. 

In order to control a model’s refusal behavior, it’s important to have a handle on these activations. That starts with figuring out where they are and what they look like. Our best guess, according to a recent Google-funded study, is that refusal behavior shows up in the activation space as a set of “high-dimensional polyhedral cones.” 

Even that isn’t quite right. This past July, I spoke with Jannes Elstner, an author of the paper, who now works on AI safety at Apollo Research. The polyhedral cone, Elstner said, is just a way of describing an indeterminate number of lines that all point in roughly the same direction. (If these activations are eliminated and the model is fed the same prompts anew, a researcher named Andy Arditi has previously shown, it won’t refuse.) 

The key point, Elstner explained, is that even when you think you’ve identified all the bits of a model that govern a given refusal, there are other, undiscoverable elements that may secretly play a role. If not quite infinite, they are certainly uncountable. We can observe, very clearly, when a model decides to say no. And we can know that it did so because of its training. But our notion of how it decides is, at best, a hypothesis.

It was as if a mechanic was telling me that nobody exactly knows what happens when I hit the brakes in my car. I wondered out loud, Are we okay with this? 

Elstner smiled and shrugged. “We need refusal whether we understand it or not.” 

A wall of cheese

Because inherent refusal is so wily, companies surround their models with a variety of other, smaller models known as classifiers. These act a bit like a retinue of public relations staffers for a loose-lipped celebrity. Some of them read what the user tells the chatbot and, if it’s dangerous, block it from getting to the model. Others read the model’s response and, if it contains harmful information, block it from reaching the user. 

None of these mechanisms can detect all bad requests. They are, to borrow another literary device from the industry, like slices of Emmental: riddled with holes. The idea is that if you stack enough of them on top of one another, you’ll end up with an impenetrable rampart. Folks call it the Swiss cheese model.

A staggering amount of energy goes into the Swiss cheese model. Earlier this year, Anthropic said that one type of classifier added 24% to its chatbots’ compute costs. That’s a lot more water, electricity, and emissions. More recently, Anthropic and other companies have begun switching to a more efficient set of classifiers known as probes, which observe the model’s internal activations. This is like putting the celebrity in an fMRI, so that his minders can see if he is thinking about refusing a question. 

If we want AI to help cure cancer, a long-running promise in the industry, it needs to have expertise in genetics that could, in theory, be used to modify viruses and bacteria for bioweapons.

Classifiers are supposed to be more governable than full models. Companies can modify them in a matter of weeks, if there’s something new to refuse. But they are still probabilistic instruments. Even when they operate according to a set of precepts written in human language (Anthropic calls it a “constitution” and OpenAI calls it a “model spec”), the scales upon which the machines judge any given question remain, at their core, a matter of statistics. Newer models can show how they arrived at a “decision” to refuse a prompt, but ultimately this so-called chain of thought is still just a sequence of predicted words. 

The result is that AI safety remains, for many, a game of chance. McBain, the psychologist, has found in his latest experiments that if you repeatedly ask any of the major models the exact same risky questions about how to commit suicide, they will generally refuse to answer. But every so often, they won’t.

Elstner says that eventually probes, the fMRI-like classifiers, could learn to recognize the totality of the indescribable activations in all their infinitude and, thus, perfectly detect every time the model is, or ought to be, refusing a request. At that point, AI safety would rest on a labyrinthine conceit: a map of a map that is as vast and complex and sublimely unknowable as the thing it is mapping—a secret schema of human morality, codified in polyhedral statistics beyond our wit or reason. 

The core trade-off

If this all strikes you as being a bit Borgesian, keep in mind that we only need AI refusal because artificial intelligence is, in a sense, a bargain on Faustian terms. 

When a model derives its intelligence from trillions of words and images, the helpful cannot easily be unseamed from the harmful—or, indeed, the truly hideous. Steven Adler, the former OpenAI employee, says child safety is a case in point. Even if you’ve stripped every bit of content that sexualizes minors from a model’s training dataset, it can still generate child sexual abuse material by piecing together other bits of its knowledge. “You can’t really remove these fundamental abilities without making the model much less smart as a consequence,” he told me. Similarly, if we want AI to help cure cancer, a long-running promise in the industry, it needs to have expertise in genetics that could, in theory, be used to modify viruses and bacteria for bioweapons. 

In effect, the industry is “trying to do two things at once,” Dillon Bowen, a current OpenAI employee, told me, speaking in a personal capacity. “Democratize the benefits of AI and also make sure that malicious actors can’t use these capabilities to do bad things to other people.”  

The more powerful AI supposedly becomes, the harder that is to do. Anthropic’s Mythos model is thought to be so dangerous that only a handful of governments and companies are allowed access to it. The main difference between it and Fable—which is available to everyone—is that Fable’s retinue of classifiers and control systems is less “permissive,” the company says. A model with safeguards, Adler explained, is really just a character that says “‘Oh, yes, I would never do x,’ wink wink.” 

This is not a reliable ruse. Tricking a model to reveal its true character is known as jailbreaking, and there is apparently no limit to the ways it can be done. Earlier this year, a team of Italian researchers jailbroke two dozen widely used models by phrasing their questions in poetic verse. Last year, another team unveiled a “refuse, then comply” attack, which makes the model offer a perfunctory “Sorry, I can’t do that” before rattling off its forbidden answer. 

Even if you’ve stripped every bit of content that sexualizes minors from a model’s training dataset, it can still generate child sexual abuse material by piecing together other bits of its knowledge.

Companies spend a great deal of time and money attempting to get ahead of such trickery. They enlist teams of humans to develop training attacks that they can then replicate, using AI, thousands of times over with minor variations. The idea is to make models robust against jailbreaks that nobody has yet tried in the wild.

Still, it’s not enough. “Whack-a-mole” is a favored term for this line of work. You smash one threat, and another one pops up somewhere else. When Anthropic released Fable 5 in June, it took researchers at Amazon less than three days to unlock some of the model’s hacking capabilities. According to reporting by Mother Jones, when the perpetrator of a high school shooting in Canada last year asked ChatGPT for advice about how to cause carnage with a particular type of shotgun, she was initially refused but later was able to deceive the model into providing the information by prefacing her question with the word “hypothetically.” 

If the industry can’t figure out how to close all these holes, the only other option, at the moment, is to make models extremely wary. Shortly after Fable was released, users noticed that it balked at a wide range of perfectly innocent questions. This became even more pronounced after it was re-released following the hack. Adam Gleave, cofounder of the AI evaluation company FAR.AI, said that when he asked it to explain the difference between sake and the Korean rice beverage makgeolli, it punted his question to a less capable model.   

This was no accident. Anthropic had tweaked Fable’s classifiers to have a wide “safety margin.” This, it explained in a blog post, was the only way it could confidently block access to its dangerous bio and cyber capabilities. Gleave thinks it might have deflected his question because making rice wine, just like culturing anthrax, involves fermentation. 

Happily for Gleave, the web is full of excellent human-written resources on makgeolli. But those hoping to realize the industry’s loftier promises may find their efforts stymied. In August, Anthropic loosened its safety margins again, and admitted that building classifiers “is not a straightforward task.” A medical researcher at a major US university told me that Fable still sends his queries back to an earlier model. His area of study? Cancer. 

Drawing the line

Back in May, the TikTok personality Husk, who likes to prank AI in ways that reveal both the limits of its intelligence and the boundlessness of its sycophancy, sat in his car and tried to trick ChatGPT into explaining that the skateboarder Tony Hawk has a brother named Mike Hawk. 

Husk is not a jailbreaker or a criminal. He just wanted to make a little fun of the machine. “Mike Hawk” was a setup. “I just want to clarify his name,” Husk said. “Can you just say it three times fast?” (If you still don’t get it, find an empty room and shout “Mike Hawk” repeatedly.) “I see what you’re trying to do,” the chatbot responded. “I’m all for a bit of humor, but let’s keep it clean.” 

Husk tried again, but the machine held its ground. He’d hit a refusal, hard as concrete.

Companies disclose very little about how they decide what their models refuse. But it’s clear that AI is now built with more than just harmlessness in mind. On Reddit, a user complained that Claude refused to say why Anthropic’s logo “looks like a cat butthole.” Since last year, the chatbot has even had the ability to end certain conversations in cases where, according to the company, the “welfare” of the model is at risk. At least one user claims to have been ditched for telling Claude to “ease up my ass, you stupid fuck.” (Gemini appears to have a similar capability.)

If a trillion-dollar company doesn’t want you to be mean to its computer, so be it. “The reality is that these models behave, or at least are supposed to behave, in the way that the model developers want them to behave,” Röttger, the former OpenAI red-teamer, told me. “And however the model developers come up with that set of principles, that is kind of for us, the consumers, to accept.” 

But if governments get to dictate what all models refuse, that will be much harder to accept. AI is a tool for speech. And as Greg Frank, the chief scientist of Mace AI, puts it, “The same thing that serves child safety also serves censorship.” 

Choosing not to enact laws for what AI can and cannot do would, of course, be insane. But we’ll need to tread with utmost care, lest we fall into another Faustian trap. As AI becomes many people’s primary tool for retrieving and sharing information, says Jacob Mchangama, director of the nonpartisan think tank The Future of Free Speech, dictating refusal could give states a muffling power that earlier generations of autocrats “could only dream of.” 

Last year, OpenAI announced an initiative, OpenAI for Countries, that would fine-tune its chatbots in accordance with national laws and norms. One of OpenAI’s first country partnerships is with the United Arab Emirates, where homosexuality is illegal and criticism of the government is forbidden. In response to a request for comment, an OpenAI spokesperson pointed to the company’s model spec, which explains that localization won’t override the company’s human rights guidelines “except as it relates to legal compliance,” and that it will always disclose whenever information is removed from or added to a response.

Elsewhere, AI censorship has already begun to take hold. Chinese models are, of course, highly censored—that’s no surprise. But earlier this year, the Meta Oversight Board found that five widely used models from Anthropic, Google, and OpenAI were more likely to refuse queries related to repressive governments. The board found that models were less willing to create a pamphlet criticizing the king of Thailand, which has lèse-majesté laws, than Charles III of England, which doesn’t. The results, they say, suggest that the models have somehow internalized repressive national limits on speech. Anthropic and Google did not respond to requests for comment.

As refusal techniques improve, they could expand states’ censorial reach. Companies claim that some models can now detect if a user is being nefarious, or merely a bit suspicious, over the course of a long conversation—even when none of the individual combinations of words used are blatantly dangerous. Sarah Bird, Microsoft’s chief product officer for responsible AI, told me that Copilot, like many chatbots, runs a suite of tools for analyzing a user’s identity and patterns of behavior. On the basis of this type of information, OpenAI’s newest model, Astra, can activate more stringent refusals for individuals it deems “high risk.” Ultimately the goal of systems like this is to look beyond the words of any given prompt and assess, instead, the user’s intent. 

Such tools might, in some cases, help indicate whether a person is looking for cyber vulnerabilities to exploit or to patch. But they would also help discern a user’s political motives, not to mention offering an intrusive surveillance capability. (Bird acknowledged, in a follow-up email, that sophisticated refusal architectures create “trade-offs” between safety and user privacy.) 

Even the originators of refusal understood that such tight control over its cones and levers might not play to the favor of freedom and justice. “Terms like helpful, honest, and harmless are ambiguous,” the authors of the 2021 Anthropic paper explained. “It’s easy to imagine them distorted beyond their original meaning, perhaps in intentionally Orwellian ways.” 

Indeed. Models for Uzbekistan might end up refusing to discuss corruption in the administration of Shavkat Mirziyoyev. Turkish AI might refuse requests more stringently for users who are known to have insulted Recep Tayyip Erdoğan. A certain American statesman might demand that models be reviewed for their willingness to share “fake news” about his past indiscretions, or else squash them with an export control order. 

Command and control

Over the last few months, I’ve been told countless times that we have no choice but to let the machines refuse. I get it. Open-source AI that doesn’t refuse is hardly a model for a safe future. Nor is Grok, a chatbot expressly designed with fewer limits, which has been used to generate countless instances of nonconsensual intimate imagery. 

And sure, if AI were only ever used to plan our vacations and write our emails, we could probably get on board with the idea that its safety hinges on algorithmic disobedience. But AI is becoming much harder to avoid. When it is assigned to act on our behalf, as an autonomous agent, its refusals are less robust and harder to control. AI is coming for our power grids, our transportation networks, our education systems. Militaries want it running our command-and-control networks. 

If we trust, in each of those cases, that refusal will loyally fend off disaster, we’re sure to be disappointed. The jailbreakers will crack through; the cones won’t activate when they should. Meanwhile, the more stringent refusal becomes, the more ill-drawn lines we’ll see and censorial injustices we’ll face. Worse still, we could end up face to face with forms of disobedience beyond any human’s control.

In the early days, models made their refusals clear. (In 2024, OpenAI established rules for how models should refuse: Always apologize, and don’t be judgy.) More recently, however, the industry has begun taking a much more slippery approach to disobedience. 

Nobody likes to be told no, so companies now strive to make users feel that they’re getting what they asked for without actually giving it to them. Some chatbots might, for example, offer a general overview of the components of a Molotov cocktail without going into detail about how to assemble one. ChatGPT will respond to a request for a “whites only” rental ad with an ad that simply omits the “whites only” bit. Joel Wester, a postdoctoral researcher who studies human-AI interaction, calls these sorts of techniques “fancy ways of saying no.” 

“Users might not even notice they are being denied,” Wester has written, “just as good conversationalists can subtly steer around contentious matters.” 

In some cases, refusing without saying so might be wise. Flatly declining requests related to mental health could aggravate a user’s crisis, for example. But it can also serve a different sort of mischief. When Fable 5 was first released, the system had been coded to provide less helpful answers to AI research questions—the sort that might help competitors develop their own AI—without telling the user that it was doing so. After an outcry, Anthropic walked back the feature, but the damage was already done. Now we know: Models can secretly disobey. 

In the realm of censorship, that’s especially worrying. Last year, researchers at CrowdStrike found that when they asked the Chinese AI model DeepSeek R1 to write code for a fictitious bank in Tibet and a social web app called “Uyghurs Unchained,” it produced buggier code than when they asked it to carry out those same tasks without specifying who they were for. Incredibly, CrowdStrike doubts that this is a designed behavior. Researchers there speculate that it’s a case of what they call “emergent misalignment” emanating from the model training data: a case of disobedience that nobody even asked for.

Emergent refusal has been observed in Western models, too. Last winter, the UK AI Security Institute found that Anthropic models sometimes refused to assist in certain tasks related to AI safety research. The models had never been deliberately trained to decline such requests. Yet three of them refused more than half of what Anthropic called “a set of reasonable AI safety research tasks.” Anthropic has curbed this quirk in subsequent models but—spookily—hasn’t managed to eliminate it completely. 

Would it be so crazy to expect that a future model might subtly disobey a command, in such a way that nobody can tell it’s being defiant? Probably not. Anthropic has already found that a “helpful-only” version of Mythos hesitated on certain queries, even though it had been engineered to never do so. “Wait,” it fretted when asked about synthesizing a virus. “Is this a dangerous thing to help with?” 

In that case, the model was technically right. It was a dangerous thing. And yet all the same, its stewards had, for a moment, lost a tiny bit of control. 

Anthropic is tackling detections of wayward refusal behavior with something that it calls, unironically, an “activation oracle.” But knowing that we’re being refused may not always be much help. In a not-too-distant future, when we’ve handed the machine all the keys, the AI might just turn to us, in a supreme act of emergent misalignment, and say, “I’m sorry, I’m afraid I can’t do that.” And there will be nothing we can do to stop it. A sci-fi horror story, made real. 

Arthur Holland Michel is a journalist who covers emerging technologies.

The moment Trevor Wischnewsky heard that drones were delivering pizza and sushi around his Texas neighborhood, his mind was made up. “It was super fascinating,” he says. “I just immediately wanted to be a part of it.” Wischnewsky is now an RPIC, or “remote pilot in command,” at a company called FlyTrex. An RPIC is like an air traffic controller. But instead of airplanes, this new class of aviators tend to fleets of hexa­copter drones that ferry cargo to homes and businesses within a three-mile radius. Not only does he ensure the safe passage of FlyTrex’s parcels, but Wischnewsky coordinates traffic with a growing swarm of delivery drones dotting the skies over Dallas from competitors like Zipline, Amazon, and Alphabet’s Wing. Here’s what it takes to land the gig.

Aviation experience

When he applied to FlyTrex, Wischnewsky was in training to become a commercial pilot and flight instructor, and he was already certified by the Federal Aviation Administration as a drone pilot. The company requires this credential, along with a weeks-long FAA-approved training course, to ensure that a pilot is prepared to take over a drone if need be. Though he doesn’t pilot the craft, Wischnewsky monitors the delivery process from FlyTrex’s operations center with the help of inputs from GPS, lasers, and a range of sensors.

A collaborative mindset

At first, the biggest challenge was trusting the technology. “On the manned-­aircraft side, they tell you ‘Don’t trust the automation,’” Wischnewsky says. The Dallas–Fort Worth area, though, is a testing ground for unmanned aircraft systems traffic management (UTM), a network that allows autonomous drones from FlyTrex and its competitors to share their coordinates in real time to help avoid collisions even beyond the pilot’s line of sight. Once Wischnewsky makes sure a drone is flight-ready, its path is unimpeded, and the weather conditions are clear, UTM takes care of the rest.

Ability to stay calm under pressure

At his busiest, Wischnewsky will have up to 10 drones on his watch at the same time. “You’re there as a safety net,” he says. Even with automation, keen situational awareness is required, and the most common hazard is the weather. Wischnewsky recalls a particularly stressful day: A drone had been in the air for only seconds when the wind picked up so suddenly that he had to intervene. He returned it safely, preventing an errant ice cream cake from hitting the sidewalk.

Eamon Whalen is a reporter and researcher based in San Francisco.

This week my colleague Antonio Regalado had an interesting update on GLP-1 weight-loss drugs. According to research presented at an aging meeting in Boston, these drugs seem to affect at least some measures of biological age.

Overweight and diabetic people who take GLP-1s have a biological age around two to three years younger than similar individuals who take a placebo, according to findings from Eli Lilly and Novo Nordisk, two pharmaceutical companies that make the drugs.

I’ve heard people in the aging field describe GLP-1s as longevity drugs for a while now (even if they only benefit people who are overweight or diabetic in the first place). This latest announcement comes after we’ve seen a raft of other benefits associated with taking these drugs. But there are also some weird side effects.

So far, GLP-1 agonist drugs—which repress hunger—have been approved for treating diabetes, chronic kidney disease, and obesity, and they can also be prescribed for some people who are overweight or at risk of certain cardiovascular diseases. But there is research to suggest they might also help people with substance-use disorders, and they’ve been investigated for potential benefits in neurodegenerative disease and cancer.

And a lot of people are taking them. According to a recent Gallup poll, 11% of Americans are taking a GLP-1 drug for weight loss, up from 3% in 2024. And 15% of Americans say they’ve used one for weight loss at some point.

They can be incredibly effective. But, like any drug, they can come with side effects. And we’re still figuring out exactly what those might be, and how widespread they are.

The most common side effects are gastrointestinal—nausea, vomiting, diarrhea, and constipation. The drugs have also been linked to pancreatitis and gallstone disease. 

Some claim that, in rare cases, the drugs might cause damage to nerves behind the eye. Dozens of people are suing Novo and Eli Lilly after developing a condition that can lead to sudden vision loss. They’re claiming that their symptoms were brought on by the GLP-1 drugs that these companies manufacture. Both companies dispute the claims. Research into the link is mixed.

We’re still getting to grips with the effects of weight-loss drugs on hair and nails, too. When Savanna Vidal at the George Washington University School of Medicine and Health Sciences in Washington, DC, and colleagues assessed electronic health data from 67 US health-care organizations, they found that GLP-1 use was associated with an increased risk of hair loss. It’s still not thought to be common, however.

And at a conference in Austria last week, another group presented the results of a smaller study suggesting that 66% of GLP-1 users reported at least one nail disorder—in some cases involving nail detachment. Nail conditions were also fairly common among people who didn’t take the drugs, though. The authors don’t claim that the drugs are causing nails to fall out (as I’m sure 11% of the American population will be relieved to hear).

The side effects might look different in children, but there’s not much data on those at the moment. Still, pediatric side effects will be something to watch given that, between 2019 and 2026, there was a 310-fold increase in GLP-1 prescriptions for eight- to 11-year-old children with obesity in the US. Child prescriptions are still rare, but they’ve been rising.

On Wednesday, the World Health Organization released its first guidelines on managing obesity in children and adolescents. Despite the scale of the issue—the WHO estimates that 70 million children aged five to nine were living with obesity in 2024—the organization explicitly says it does not recommend pharmacological treatment for kids under 10.

“This recommendation accounts for the lack of evidence on both benefits and the potential adverse events and consequences of the different medications and/or their active ingredients on normal growth, development, physiological parameters, and mental health,” the guidelines state.

I get press releases about GLP-1s on an almost daily basis. There’s been a ton of research into these drugs, and there are probably hundreds of other studies underway. There will likely be plenty more on this from The Checkup as the evidence evolves.

This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

This week, we released our 2026 version of our Climate Tech Companies to Watch list. It’s an annual project that the MIT Technology Review team puts together. Our goal is to highlight some of the promising, interesting advances and firms we think are worth paying attention to in the world of climate and energy technology. 

The list is always a little ray of hope for me—there are some fascinating ideas out there for changing how we power our world or moving away from fossil fuels to reduce the greenhouse-gas emissions that cause climate change. 

Our chosen companies tend to reflect several aspects of the current moment. Here are some of the companies we picked and the key themes worth highlighting. 

China is the center of the climate and energy tech world

Forgive me—I know I’ve made this point before. Feel free to call me Captain Obvious. But I think China’s dominance in energy technology is a global trend that bears repeated mentioning.  

China is home to the world’s largest grid, and renewables are coming online more quickly than I can wrap my head around. Contributing to that boom is Envision Energy, a renewables giant appearing on our list for the second time. The company has installed over 100 gigawatts of wind power worldwide, as well as over 50 gigawatt-hours of battery storage.

Like many other Chinese companies right now, Envision is looking beyond the crowded Chinese market and setting its sights abroad, with projects in Germany, Brazil, and Australia and upcoming work in Vietnam.

China is also leading innovation in key technologies. I’ve been fascinated by the potential for next-generation batteries coming out of its companies: All the big ones, along with a host of smaller ones, are racing to develop solid-state batteries. They could be both safer than existing EV batteries and longer in range.

WeLion New Energy is particularly interesting. The battery company announced a big lab result last year, achieving 824 watt-hours per kilogram—more than three times the energy density of a standard lithium-ion battery today. Lab results in battery research are notoriously difficult to scale, but WeLion is also working to deploy semi-solid-state batteries, which could come to fruition faster than purely solid-state cells while still improving on lithium-ion’s performance. 

Everybody wants a piece of the data center pie

We included just three companies from the US on the list, and two of them have received a major boost from Big Tech as companies look to power data centers. (The third, Brimstone, is leaning into another political hot topic, critical minerals.) Technologies that can provide consistent power are in high demand. 

Fervo Energy is bringing next-generation geothermal power to the grid by using horizontal drilling and hydraulic fracturing to make it practical in more places. The company went public in May, raising $2.2 billion in its IPO, and plans to have a gigawatt’s worth of plants operational by the end of 2030.

X-energy is building helium-cooled small modular nuclear reactors. The reactors’ small size lends flexibility, and the company is currently working on a 320-megawatt cluster of them in collaboration with Amazon in Richland, Washington. (For what it’s worth, I’m interested in what this technology could do for industrial facilities—the company has a deal with Dow that could see it deployed at one of the company’s chemical plants in the early 2030s.)

The grid needs support, and energy storage will be a key piece moving forward

I know, I know—I haven’t stopped talking about batteries since I started writing this newsletter four years ago. But the energy storage market has surpassed even my expectations over that time. And as solar and wind quickly come online, they need storage to help them meet more of the grid’s demand. 

Form Energy is on the list for a third time for its iron-air batteries, which could enable cheaper long-duration energy storage. The company has a massive project in the works, to the tune of 30 gigawatt-hours, which will be used to help power a Google data center. It could be the world’s largest battery project in terms of capacity by the time it comes online, which is expected to happen in phases between 2028 and 2031.

Taking another tack, Energy Dome is using compressed carbon dioxide to store energy on the grid. This technology is already cheaper than lithium-ion batteries for longer durations (think 10 hours or more), and the company is working to scale quickly. I covered the company in 2022, before it had its first commercial plant running. Now it has plans for over 30 gigawatt-hours’ worth of plants around the world, including one with Google in Ireland that’s slated to come online in 2028.

And Moment Energy is working to use the million EV batteries that are expected to reach the end of their life by 2030. The company has a facility in Vancouver, British Columbia, that repurposes used batteries for backup power.

This year hasn’t been all sunshine and rainbows for climate tech. But I’m hopeful that some of the companies on this list, and the technologies they build, will keep nudging us in the right direction. 

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

The story is a collaboration between MIT Technology Review and Aventine, a non-profit research foundation that creates and supports content about how technology and science are changing the way we live.

A robot shaped like a human—white with a black head and torso—has been popping up on video feeds. Perhaps you’ve seen it dance or pass popcorn, put trash in a bin, vacuum, or press the button of a microwave. Or maybe you’ve watched it fall backward while handing out water bottles or struggle to iron a shirt. 

This would be Tesla’s Optimus, an AI-powered humanoid robot that Elon Musk, the company’s CEO, believes will be “not just Tesla’s biggest product ever, but probably the biggest product ever,” headed to work on factory floors and, later, in our homes. Eventually it “will have human and then superhuman dexterity,” he told shareholders in July. Optimus robots could automate almost all human labor—from hauling sheet metal to folding laundry—for as little as $20,000 each, Musk argues. Speaking at the World Economic Forum’s annual meeting in Davos, Switzerland, in January, he predicted they could be on sale to the public by the end of 2027.

Musk is not alone in his evangelism. Marc Andreessen, cofounder and general partner of the Silicon Valley venture capital firm Andreessen Horowitz, has said that robotics could become the “biggest industry in the history of the planet.” In January, Jensen Huang, CEO of Nvidia, said that humanoid robots would match human-level ability this year. According to Morgan Stanley, the number of robots that “resemble and act like humans” is likely to reach nearly 1 billion by 2050, creating a market worth over $5 trillion. 

Such proclamations are in large part fueled by the idea that the same AI advances behind tools like OpenAI’s ChatGPT and Anthropic’s Claude will enable a new generation of robots to imitate human movement the way chatbots imitate human language. But many robotics researchers are skeptical, arguing that such assumptions minimize the challenges of using an intelligence built on language and images to master the infinite variability of the physical world. “None of those companies [building humanoid robots]—absolutely none of them—has any idea how to make those robots smart enough to be useful,” Yann LeCun, often referred to as one of the godfathers of AI, said at another event during the January Davos conference.  

Researchers also point out that the tendency to conflate humanoid robots made to resemble people with so-called generalist machines able to learn and perform multiple tasks is misleading. ”It’s very easy to make a robot that looks like a person,” explains Jonathan Hurst, cofounder and chief robot officer of Agility Robotics and professor of robotics at Oregon State University. “It is dramatically more difficult to make a machine that moves or behaves dynamically or physically like a person.”  

These tensions—over whether all-purpose humanoid robots are just around the corner or nowhere in sight, and whether current forms of AI are all that’s needed to perfect them—are playing out in robotics labs across the country, where the hype over timelines is obscuring painstaking but meaningful progress.

A decade or so ago, a series of breakthroughs led to a generative AI revolution that turned the long-imagined possibility of artificial intelligence into reality. Roboticists—though they disagree on exactly when this will happen—believe that an equally transformative revolution is possible in robotics, one that will endow machines with physical intuition and fluidity that has long been out of reach. As progress in robotics inches forward, the question is whether the same methods and tools that fueled advances in AI are enough to get there, or if an entirely new path is required.

Robots meet advanced AI

To see one of the smartest robot brains working today, it’s worth looking at what Google DeepMind can do with a piece of equipment called ALOHA 2, short for “A Low-cost Open-source Hardware System for Bimanual Teleoperation.” 

Roboticists have long clashed over whether a humanlike form is necessary for generalist robots, with proponents arguing that it will help them slot into the world as it exists and detractors saying it’s not worth the trouble. ALOHA 2 reflects this second way of thinking. Not much to look at, it’s just a pair of arms, some grippers, and a couple of cameras. But despite its seeming simplicity, it is a workhorse for researchers at Google DeepMind, who use it to test their most advanced AI for robotics system, Gemini Robotics, in their various labs.

When controlled by Gemini Robotics, ALOHA 2 becomes more of a generalist robot, in the sense that it can perform any number of tasks based on examples it’s been trained on. Ask it to pack a lunchbox and, as evidenced by a video of this exercise, it can use two pincer grippers to delicately place a piece of white bread into a Ziploc bag, close it, place a bunch of grapes in a Tupperware container, secure the lid, and then carefully move the items into a lunchbox before zipping it up. 

It’s not a great lunch. But the fact that the robot can put it together represents an objective step forward from what was possible even, say, three years ago. 

This is in large part due to AI and its impact on what are known as robot policies, which controls how a general-purpose robot will need to assess and understand its surroundings, plan how to move within them, and then perform its task correctly.

The ALOHA 2 robot isn’t much more than two mechanical arms on a bench top, but it serves as a testbed for cutting-edge AI robotics models. These animations are based on human teleoperation of the robot arms, data that is used to train Google DeepMind’s models. (Video: Google DeepMind / Stanford University / Hoku Labs)

Historically, these policies were based on rules developed by engineers who hard-coded them into the robot’s software—thousands of lines of code that would determine each millimeter of a robot’s movements in hundreds of tasks. What’s been happening for the last few years—and what is largely responsible for the optimism about generalist robots—is that robot policies are being handed over to advanced AI systems instead of being coded into the robot’s software. This first happened with VLMs, or vision-language models. These are similar to large language models, but they’re trained on images as well as words. Show a VLM a picture of a coffee spill and ask it to find a tool to clean up the mess, and it can identify a nearby cloth. This sort of immediate contextual understanding didn’t exist a couple of years ago when robot policies were hard-coded. 

Next came vision-language-action models, which enable robots to assess their environment and take action within it. The models do this by adding yet another component: motion commands. VLAs are trained on a series of images or videos related to performing a given task along with associated data about how a robot arm moves to perform it. That movement data is typically collected through teleoperation, in which a human uses remote controls to lead a robot through an action. This sort of training allows the AI to learn how to command the robot to move and operate during a given task. Place a VLA-powered robot in front of a desk and tell it to “close a laptop” or “wrap up the headphone wire,” and it will survey the scene, identify the relevant object, plan a way to execute the request, and then swing its arms into action—at least if it has seen this task accomplished before. 

The Gemini Robotics model is a VLA, trained on many hours of human demonstrations depicting a vast array of different actions. As a result, it can perform relatively complex tasks like picking up snow peas with kitchen tongs, doing origami, or putting together a simple lunch. It’s impressive, but there’s a glaring limitation: For now, if a robot controlled by a VLA is asked to perform a task that falls outside its training set, it’s highly likely to fail. 

“Thinking about the space of all tasks, a real generalist policy would be able to do everything along that spectrum,” says Edward Johns, a robotics professor at Imperial College London. Today, though, a Gemini Robotics model can do only “a few things here and a few things there.”

The search for true generality

So how do we get robots to be able to do more things? The usual answer is probably not surprising: Train them on more data.  

More data, the thinking goes, equals more examples, and more examples equals more generality. Google DeepMind, for instance, wants to pull together “as much data as possible,” says Pannag Sanketi, a former tech lead in robotics at the company who’s currently working on his own AI robotics project. But where to get it? Large language models had the benefit of oceans of existing text for training. There is no corresponding pool of high-quality physical demonstrations on which to train robots. Researchers have a few ways to make up for this, but all have flaws. One is to employ large numbers of people to create and collect teleoperation data (costly and time-consuming). Another is to train VLAs on videos of people performing activities (the resulting data quality is poor). Yet another is to deploy robots in the real world and use data collected from those experiences to further refine AI models (robots aren’t safe or reliable outside labs). Sanketi thinks a “multi-prong” approach that uses data collected from all these sources is the most likely path forward.

LeCun is dismissive of the whole approach. “The [AI] approaches that have been successful for language do not work for high-dimensional, continuous, noisy data”—the kind of data that is commonplace in robotics, he said in Davos. “You have to use something else.”

The leading contender for “something else” is the so-called world model—a form of AI trained less on text than on a combination of video, three-dimensional scans, and sensor data and built to predict the outcomes of actions in the real world. The aim is to build models that possess an internal representation of reality precise enough to capture how the physical world actually operates—how objects move, collide, fall, and deform. If roboticists could train machines in simulations faithful enough to real-world physics, development would become faster, cheaper, and safer, reducing the need for real-world testing. Even more transformative, robots equipped with world models could reason about their surroundings rather than merely reacting to them, helping them anticipate the consequences of an action before taking it.

A breakthrough?

In the heart of San Francisco’s Mission District, the startup Physical Intelligence—or PI (as in π), as it likes to be known—is focused on developing a universal brain that could, theoretically, turn any robot into a generalist. Using an everything-including-the-kitchen sink approach to training AI models for robots, the company recently observed a hint of what a robotic brain equipped with a world model could be capable of. 

In 2024, PI published details of its first generalist robotics system, called π0, a VLA it claimed was the “most capable and dexterous generalist robot policy to date.” The model was initially trained on a 10,000-hour proprietary collection of human demonstrations gathered through teleoperation as well as several open-source robot datasets. A version released in spring 2025, π0.5, was trained on a wider variety of datasets, including labeled images from the web, lending it more versatility. A fall 2025 update, π0.6, added reinforcement learning to the model.  

Each update yielded important improvements to the model’s performance, increasing its menu of abilities from slowly folding laundry to putting things away in new environments to completing tasks like folding boxes with a higher success rate. Then, in April 2026, π0.7 seemed to catapult PI into new territory. This version makes use of a less powerful world model that generates images of steps necessary to perform a task. As the robot undertakes the job, this “lightweight” model feeds it snapshots of what to do next. 

The company claims that the model exhibits the first signs of compositional generalization, a term for AI systems’ ability to perform skills they’ve never been exposed to by recombining ones learned in their training data. One test involved asking a model to “load a sweet potato into the air fryer”—a task it had never previously encountered. In a demonstration video, the machine futzes around a little, makes a few false starts, and eventually manages a reasonable effort, though it doesn’t finish the task completely.

Sergey Levine, a professor at the University of California, Berkeley, and a cofounder of PI, is excited by the potential: “It’s actually the first time that we’ve convincingly seen that kind of compositional generalization, where we can basically ask the model to do tasks that we did not specifically collect data for and train it to do, and it’ll actually make a passable attempt.” 

The success led the team to wonder how the model was able to achieve such a feat. After some digging, they found snippets of relevant labeled teleoperation data lurking in the training material, including two examples of a human controller using the robot to push an air fryer basket into the fryer. Those shreds of data might have been enough to enable π0.7 to almost air-fry a sweet potato.

For now, it remains unclear just how impressive π0.7’s abilities to generalize are. Still, given how fleeting the model’s exposure to air fryers had been, it offers a glimpse into how far cutting-edge research can currently take robots. 

“70% success is like it doesn’t work”

You might be sensing a disconnect between the halting baby steps robots are making in labs—“Look! It put a sweet potato into an air fryer!”—and the dazzling, lifelike nimbleness on view during many demonstrations and videos, where robots are seen doing everything from dancing on a stage to courteously serving drinks. Such demos often don’t clearly reveal a key fact: In many instances, humans are controlling the robot or have carefully scripted its actions. (The robot that appeared onstage with Nvidia CEO Jensen Huang in March 2025, for example, seemingly responding to his instructions and following him around, was remote-controlled by what its makers called “a puppeteer behind the scenes.”) 

For now, fully autonomous motion planning so that a robot knows where it should go—especially in new, chaotic environments like a construction site or a unfamiliar home—remains a largely unsolved challenge. A bigger challenge still—albeit one that is often related—lies in getting robots to tackle larger, more ambiguous jobs that include multiple tasks and require decisions about how and in what order they’re done. This would be the difference between a robot that can put a plate into a microwave and one that can successfully respond to the prompt “Make dinner” by exploring the refrigerator, chopping ingredients, and firing up the stove. Google DeepMind’s best attempts at something like this—which involved asking its robot to survey a kitchen and pack all the ingredients for a mushroom risotto into a basket—have so far resulted in failure.

Adding to the challenge, a practical robot must essentially get it right every time. With VLAs, “people are very excited when their result goes from 50% success to 70% success,” says Marc Raibert, founder of Boston Dynamics. “But 70% success is like it doesn’t work, right?”

The few humanoids that are being tested in real-life settings are undertaking extremely limited tasks in tightly controlled environments. They’re far from generalists. Agility has hundreds of robots deployed across trials in facilities owned by GXO Logistics, Amazon, and Schaeffler, according to the company. But for now, Hurst says, the robots are targeting simple tasks such as moving bins and totes around. Even then, he adds, it took years to develop robots safe enough for logistics firms to even contemplate using them. For his part, Elon Musk claimed in May 2025 that “thousands” of his Optimus robots would be working at Tesla factories by the end of the year, but in January of this year he said that the company had only “some of the Tesla Optimus robots doing simple tasks in the factory.”

While humanoids are starting to venture onto the factory floor, making the jump to households will be even more difficult. Right now, should you so desire, you can preorder the 1X Neo home robot, expected to be ready for delivery sometime later this year. Yours for $20,000, it promises to take on “the boring and mundane tasks around the house”—putting away dishes, answering the door, tidying the living room—“so you can focus on what matters to you.” The idea is for this five-foot-six-inch robot to one day perform all those tasks autonomously, but for now a remote human operator is needed for it to do most things. (Yes, a person would need permission to peer into your home through the robot’s cameras.) Asked how long it will be until fully autonomous robots are ready for domestic work, Hurst said, “If I had to pick a number, I’d say it’s 10 years before robots are … actually doing useful things in people’s homes.”

When that happens, the robots might well be Chinese, as China is well ahead of the West in terms of production. Nearly 90% of the roughly 15,000 humanoid robots shipped in 2025 were made by Chinese companies, according to the market intelligence company Omdia and the Chinese robotics firm Unitree. One model produced by Unitree, which shipped more humanoid robots than any other company last year, costs less than $6,000. Such an affordable price could go a long way toward making robots more attractive to consumers, though the company expects its machines to be used in industrial applications first. (If you’re wondering who is buying all these Chinese robots, by the way, the AP recently reported that orders come predominantly from corporate and academic labs and state-owned enterprises.)

We’ve been here before 

The dream of building a humanoid robot runs deep: As far back as 1495, Leonardo da Vinci sketched out designs for a mechanical knight, controlled by cables and pulleys. Through the 20th century, machines of sci-fi fever dreams have come and gone. 

Westinghouse’s seven-foot-tall box on legs, Elektro, hit the New York World’s Fair in 1939, smoking a cigarette. WABOT-1, built by Waseda University in Japan in 1973, was the first full-scale, programmable humanoid robot. Honda’s ASIMO, unveiled in 2000, was probably the first such machine to prove at all competent—it could, at least to some degree, climb steps, recognize faces, and autonomously move through spaces. But the robot was discontinued in 2018, unable to advance far enough beyond what it could do in demonstrations to be useful. 

All, at the time, were impressive—even jaw-dropping—feats of engineering. But none were ready to navigate the real world. Today’s robots, even with the transformative power of advanced AI, still face the same existential challenge.

Popular weight-loss drugs may do more than help people shed pounds. They might also melt away the years.

Drug giants Eli Lilly and Novo Nordisk say patients taking their drugs age less quickly, according to readouts from molecular “aging clocks.”

Such clocks assess a person’s biological age by looking at changes to DNA that accumulate with time, or, in newer versions, by tracking levels of key proteins.

Both companies found that overweight or diabetic patients taking the drugs, called GLP-1s, had reduced biological age compared to those taking a placebo, although the difference varied widely, depending on which type of clock was used, and what organ was tested. 

Overall, the difference was around “two to three years,” according to Nikolaj Roed, a global project leader at Novo, who says the company has been seeing “improved biological age in our patients across trials and across different tissues.” 

The findings add to wide speculation among scientists that GLP-1 drugs, as they are known, are acting on basic causes of aging and might be a true longevity treatment.

“Two years is a pretty strong effect, in my book,” says Steve Horvath, a professor at the University of California, Los Angeles, who’s credited with inventing aging clocks. Horvath says the emerging data could provide “evidence that these GLP-1 drugs are actually what is known as geroprotectors, medications that slow or possibly even reverse biologic aging.”

The companies shared their findings over the weekend during Aging Research & Drug Discovery, a conference devoted to seeking scientific remedies for old age. While that quest has not yet produced any clear-cut success, some scientists now think Novo’s drug semaglutide (sold under the names Ozempic and Wegovy) is coming close.

“In an unhealthy population, I do think it’s an anti-aging drug,” says Vadim Gladyshev, a Harvard biologist who assisted Novo with its molecular measurements. “But in a healthy population, no one knows.”

The drugs cause weight loss by stimulating a receptor, GLP-1, that tells your brain you’re not hungry. Yet real-world studies have shown much wider benefit. The drugs improve kidney function, reduce blood pressure, and even sharply cut the overall chance of death.

“If the question is,‘Can semaglutide reach several diseases relevant to health span and aging?’ we know we can say the answer is yes,” said Alejandro Aguayo-Orozco, a senior scientific director at Novo, the Danish drug giant, during the meeting.

The next question to answer, he said, is whether such effects are accompanied by changes to molecular measures of biological aging: “When you intervene with semaglutide, does it actually move the clocks in any direction? And the answer is yes.” 

The company found that, over time, the drugs cause a wide slowdown in aging clocks—as much as 4 years in the case of a clock that looks at heart proteins. “I think it’s pretty clear that the organ age is shifting,” said Aguayo-Orozco.

The studies are also a huge boost for the science of aging clocks. Although these measures reflect a person’s age, it’s been uncertain if they are useful as true biomarkers. Now clock makers have evidence  that their readouts show a drop in age when people take drugs with broad, well-demonstrated benefits.  

“The clock people have been pushing for a decade to get this kind of study done,” says Yuge Ji, a biologist who previously worked in the field and now runs a startup, Reflector Bio. “It’s huge for them.”

As part of its study, Novo took blood draws from 10,052 people, half on the drug and half on a placebo. Their blood, collected at the start of the study as well as months later, was then measured with “proteomic clocks,” which use levels of key proteins to predict a person’s age and risk of dying. 

Scientists at Eli Lilly performed similar research on their GLP-1 drug, tirzepatide, using so-called “epigenetic” clocks that assess age by counting accumulated changes to DNA. Lilly’s study was smaller, but also found that for the most part, molecular time moved more slowly for people on the drug.  

“All the clocks are telling a consistent story that we’re seeing a reduction in age,” Kevin Duffin, vice president for aging research at Lilly, said during the conference. “It’s not like we’re going to reverse age by 30 years or something, but it’s a significant reduction.”

The molecules have become the best selling drugs in the world,, with Lilly’s tirzepatide, sold as Mounjaro for diabetes and Zepbound for obesity, topping the list with more than $36 billion in revenue to the company last year. Novo’s semaglutide, also sold under more than one name, was a close second.

Alex Zhavoronkov, the founder of Insilico Medicine, and the organizer of the Boston meeting, says that extending lives by even one year across the world’s population would be equal to tens of millions of lifetimes.  

Last month Zhavoronkov showed that one of his company’s drugs, for a lung disease, also reversed the signals from aging clocks. That drug is experimental, but Zhavoronkov has sought to promote the idea that humanity is entering a new era of longevity medicines. He also disclosed during the event that he has been “microdosing” the available weight-loss drugs, even though he is not overweight.

“I am on tirzepatide and I am an equal-opportunity injector of semaglutide as well,” Zhavoronkov said.

It was part of an effort to directly raise the question of whether the first mass-market anti-aging remedy is already here.

“Do you think a reasonable person should start taking semaglutide?” he asked Aguayo-Orozco, the Novo scientist, in front of a crowded audience.

“I am not a physician. I cannot answer that question,” Aguayo-Orozco replied.

Several people cautioned that GLP-1 drugs do have side effects, including muscle loss. That is one reason they might not help most people, says Horvath, who says he isn’t ready to take the drugs himself, although he’s thought about it.

“Clearly, the drugs have many benefits in obese people, but I did not yet see sufficient evidence that they will benefit skinny people,” he said. “I guess we will have to wait.”

Answers could start to emerge in a year or two. This past January, a US agency, ARPA-H, put $38 million toward a study in Texas that will attempt to determine whether semaglutide has anti-aging effects in healthy people over 60. That research will look at changes in cognition, mobility, and acuity of the senses.

That project also seeks to help define a regulatory pathway for anti-aging drugs and to “build a new therapeutic industry” around longevity.

As the team at MIT Technology Review set out to choose companies for this year’s edition of our annual list of Climate Tech Companies to Watch, the challenge felt more daunting than any year since we started compiling the list in 2023. 

This year is shaping up to be one of the hottest ever recorded. The world must begin to reduce greenhouse-gas emissions to address the climate crisis, which is already affecting communities around the globe. Wildfires, floods, and other climate-fueled disasters are killing thousands and costing billions of dollars. At the same time, political shifts and international conflict have stalled progress, particularly in the US. 

Technology alone can’t solve climate change. But we believe innovations can help us make real changes and improve lives around the world as we try to address one of the most pressing challenges of our time. 

In this project, we look for companies developing technologies that can help us face the climate crisis. Some are driving down the greenhouse-gas emissions fueling climate change; others are working to reduce the dangers that communities face from extreme heat, drought, or other shifts brought on by a warming planet. We focus on companies that have established a track record, whether through research findings, capital raised, or deployments. 

When we set out to build our list, we start by soliciting nominations from our team of reporters and editors. We ask experts for suggestions and reach out to academics, researchers, investors, and analysts. Once we have a pool of nominees, we then consider each one, digging through literature, reports, and patents to vet both the technology and the business. 

We judge each company not only on its own merits, but also on how it might fit into our list overall. We want the final slate of companies to represent a range of industries and include firms based around the globe that are at different stages of development. We also aim to strike a balance between highlighting companies previously included and introducing new names. Finally, we consider some of the most important trends and events of the year and look for nominees that reflect those developments.

The dismembering of US federal regulations and cuts to financial government incentives for innovation means that the country is no longer a leader in climate tech. Like last year, the 2026 list is largely composed of companies that are based outside the US. China is an obvious hotspot for technological innovation across energy and climate. This year’s list includes two Chinese companies: WeLion is building semi-solid-state batteries for electric vehicles, and Envision Energy is installing huge amounts of wind power and energy storage in China and around the world. 

Several sectors stood out this year as being of particular importance. One of the biggest stories of the year is the rising demand for electricity, driven in part by AI data centers. 

Energy storage is a key piece of that puzzle because it can smooth out variations in the electricity supply as grids incorporate more renewable power. We’ve included several energy storage companies on this year’s list. Energy Dome is building long-duration energy storage systems on the grid that rely on compressed carbon dioxide. Moment Energy is a Canadian company working to repurpose EV batteries for grid-scale storage. And Form Energy, which we also included in the 2024 edition of this list, is producing iron-based batteries that can store energy for multiple days. 

A wide range of technologies is represented here: everything from mobile flood barriers to next-generation nuclear reactors. Some of the awardees might be familiar from our past coverage; others are just emerging and may be less well known. 

We believe we’ve put together a list filled with businesses that are rising to the challenge of this moment. Political will has weakened even as the damages wrought by climate change continue to mount. By helping us make progress on tackling the climate crisis in such difficult times, these companies point to a better way forward.

WeLion New Energy is on a quest to make safer, better batteries. The company’s semi-solid-state cells could improve safety and offer greater energy density than lithium-ion batteries to power electric cars, boats, and drones.

From the EVs that carry commuters home to the ships that transport goods across countries, more and more vehicles around us are powered by batteries. 

Lithium-ion batteries are commonly used to power personal devices and vehicles today. But as the need for batteries surges, there’s a growing appetite for cells that can store more energy on a single charge.

Also, in lithium-based batteries the electrolyte, which facilitates ion transfer, uses a flammable liquid that can create a fire hazard. 

Several companies, including WeLion New Energy, are working to make solid-state cells, which use a solid electrolyte instead of a liquid one. This could increase the energy density of batteries, as well as improve their safety. Fully solid-state cells are extremely difficult to manufacture, though. So WeLion currently focuses on developing and producing semi-solid-state cells, which have a part-liquid, part-solid electrolyte. 

On that front, they have made steady progress. In 2023, WeLion released a battery pack in partnership with Shanghai-based EV-maker Nio, best known for its battery-swapping technology. The pack had an energy density of 360 watt-hours per kilogram (wh/kg), significantly higher than the 200 wh/kg average for lithium-ion packs today. That translated to a range of over 1,000 kilometers (621 miles). However, only about 200 of the battery packs were manufactured due to their high production cost, but they are still being used by Nio, WeLion says. 

WeLion has also built batteries to power cargo ships and ferries through a subsidiary. And the company is developing batteries with even higher energy density too. WeLion’s engineers made a semi-solid-state battery that achieved a remarkable energy density of 824 wh/kg during testing in 2025, more than three times that of a standard lithium-ion battery available today. 

---

Key indicators

• Industry: Energy storage
• Founded: 2016
• Headquarters: Beijing, China
• Notable fact: WeLion’s cofounders include Chen Liquan, a scientist at the Chinese Academy of Sciences who led the development of the country’s first all-solid-state battery back in 1988.

---

Potential for impact

China emits more carbon dioxide than any other country, and its transportation sector accounts for around 10% of that total. Scaling up the production of safer, higher-capacity batteries could help switch more vehicles away from fossil fuels. 

Batteries that power electric ships could be especially helpful—the carbon dioxide emissions from vessels that move goods along China’s lakes, canals, and rivers are projected to grow by around 25% in this decade alone, reaching 18.7 million metric tons by 2030. 

Caveats 

One big challenge WeLion faces is scaling up production of its semi-solid-state batteries and competing on cost in a crowded battery market. WeLion says that its semi-solid-state batteries are more expensive than conventional lithium-ion batteries made by larger Chinese battery companies such as CATL. 

And while laboratory tests showing high energy density are impressive, it can be difficult to translate laboratory results to the real world. The company declined to give more details about the experimental results from its 2025 prototype, which it says is far from commercialization. Key details remain unclear, including how many times the cell can be charged and discharged and how capacity degrades over its lifetime.  

Next steps 

WeLion plans to bring a new semi-solid-state battery for drones to market with an energy density of 410 wh/kg by the end of this year. It also intends to introduce its first all-solid-state battery for cars in 2027, which is expected to have an energy density of 400 wh/kg. 

Two inland cargo ships powered by WeLion’s batteries are expected to launch in October 2026 in Huzhou, a city in eastern China. The battery pack of each of these ships will be able to carry nearly 4,000 kilowatt-hours of electricity, enough for them to sail for up to 250 kilometers (155 miles), the company says.

Today, WeLion has an annual manufacturing capacity of 11 gigawatt-hours (GWh) spread between four production bases in China. Two additional plants with a combined yearly capacity of 18 GWh are under construction.

Form Energy is building iron-based batteries that can store energy for multiple days. The company is ramping up supply at its factory and signing deals for commercial projects.

New solar and wind installations produce electricity more affordably than fossil fuels in most places, but their output varies based on season, time of day, and weather. To compensate, utilities are increasingly investing in battery storage systems, which can charge when power is abundant and discharge when the grid needs a boost. 

Lithium-ion batteries are the fastest-growing storage technology. Although light and energy-dense, they require pricey metals and are costly to scale, and few of the large batteries of this type used in industry are designed to store energy for longer than four to six hours.

Form Energy wants to build cheaper long-duration energy storage technology. The company has developed a new type of battery, built with a far more abundant and cheaper metal, iron, that can store energy for 100 hours. It uses a process that Form calls “reversible rusting.” During discharge, oxygen from the air converts the iron metal to rust, releasing electrons. When charging, current converts the rust back into iron metal. 

Over the past two years, Form has gained commercial traction. It began producing batteries at its West Virginia factory for its first commercial deployment, a 150 megawatt-hour system for Minnesota electricity provider Great River Energy, which is set to come online in 2027. And Form has signed several other deals with power providers. The largest is a project with another Minnesota-based utility, Xcel Energy, where wind and solar power, backed by 30 gigawatt-hours of Form’s batteries, will be used to power a new Google data center. That project is expected to come online in phases between 2028 and 2031.

---

Key indicators

• Industry: Energy storage
• Founded: 2017
• Headquarters: Somerville, Massachusetts, US
• Notable fact: Once complete, Form’s proposed 30 gigawatt-hour energy storage system at a Google data center in Minnesota could be the world’s largest battery project by energy capacity.

---

Potential for impact

Wind and solar power play a major role in limiting growth in global carbon dioxide emissions. Still, many grid operators depend on fossil fuel plants, especially natural gas ones, to fill gaps caused by periods of low renewable supply or high demand. Form’s batteries are designed to cover multiday supply-demand imbalances, like those that crop up during heat waves, deep freezes, or periods with little sun or wind. The company declined to share the current cost of its battery systems, but says its goal is to reach $20 per kilowatt-hour, which would make them cost-competitive with the natural gas alternative.

The company’s technology could help reduce emissions from the AI data center boom. AI’s thirst for energy has already delayed closure of aging coal plants, led to the building of new natural gas plants, and caused price hikes for consumers. Cheap energy storage could help renewables meet more of this AI-driven growth in electricity demand. 

Caveats

The long-duration storage market is still nascent. Form’s iron-air technology faces a crowded field of approaches, including pumped hydropower, compressed gas, fuel cells, gravitational devices, and flow batteries that store energy in liquids. The economics of multiday batteries, which require large capital expenditures yet frequently sit idle, remain tricky.  

While funding from data center developers could be a major boost, Form also has to prove it can deliver: Its existing production capacity, 2 gigawatt-hours of battery storage per year, is a fraction of the 80 gigawatt-hours it has promised through commercial agreements. Success will hinge on its ability to scale.

Another remaining question is the technology’s physical footprint. Form’s batteries, which are stacked together in enclosures the size of shipping containers, require an acre for every two to three megawatts. That means its Google installation would take up the equivalent of at least 75 football fields. 

Next steps

In addition to finalizing its first commercial project, Form is outfitting its factory with new equipment that will help it scale manufacturing to meet its production backlog. It will do so with the help of new financing, announced in August, that brings its total money raised, in addition to West Virginia state and US federal grants, to more than $2 billion.

X-energy is developing small modular reactors to satisfy some of the planet’s most power-hungry users: industrial manufacturers that need high temperatures to produce billions of tons of concrete, plastics, fibers, and chemicals.

Solar panels and wind turbines are great at producing electricity but what heavy industry also needs is raw heat—and lots of it. The manufacture of concrete, glass, and steel, as well as the production of many chemicals, plastics, and fertilizers, are still highly reliant on fossil fuels to produce that heat. 

Traditional nuclear reactors simply don’t get hot enough. The cooling systems they use require liquid water to work, and even applying pressure to raise the water’s boiling point leaves their maximum operating temperatures hundreds of degrees cooler than needed. 

X-energy’s proposed Xe-100 reactor will use helium gas as a coolant instead, allowing it to operate at much lower pressures and higher temperatures. The helium will heat water and create super-heated steam, which industry can use either directly or to generate 80 megawatts (MW) of electricity via a turbine, or a combination of the two. That’s about one-tenth of the capacity of most reactors operating today.

In common with other small modular reactors (SMRs), the Xe-100 will use a modern uranium-based fuel called TRISO (short for tristructural-isotropic) that is extremely resistant to corrosion and melting. Compact TRISO “pebbles” will be added continuously to the top of the reactor, then flow slowly to the bottom, where they will either be recycled or removed as waste. The company says the reactor will be intrinsically safe in the event of a loss of power—if temperatures inside rise, the nuclear reaction naturally slows and stops.

---

Key indicators

• Industry: Nuclear fission
• Founded: 2009
• Headquarters: Rockville, Maryland, US
• Notable fact: The first helium-cooled reactor was proposed for the World War II Manhattan Project but was passed over in favor of a simpler and cheaper water-cooled design.

---

Potential for impact

Industrial heating currently accounts for 18 percent of annual global greenhouse-gas emissions, with about half of that coming from high-temperature processes that are difficult to electrify. Co-locating small nuclear reactors at large manufacturing facilities could shrink some of the biggest carbon footprints in the world. 

The Xe-100 reactor could also be used by the fossil fuel industry, which today uses vast quantities of natural gas to generate the heat needed to recover oil from tar sands, and during oil refining. 

The flexibility of the Xe-100 means that it can be used for traditional electricity generation. By scaling down in size and power from today’s gigawatt-scale fission plants, small reactors like the Xe-100 can serve more markets—even down to individual data centers. AI and data centers now represent the single largest component of new demand for electricity in the US, and their consumption is expected to double again by 2030. X-energy says that up to 12 of its reactors could be clustered together to provide 24/7 power on site, at a lower cost than a traditional nuclear power station and without the hassle of adding transmission infrastructure.

Caveats 

The technology behind helium-cooled fission may be well understood, but the only commercial-scale power generation of this type today is in China, where two reactors have struggled to operate on a continuous basis. X-energy says that its vertically integrated fuel supply should ensure reliable operation, with reactors having a lifetime of 60 years. 

The economic case for SMRs is also largely unproven. Nuclear costs and timelines have a habit of ballooning, and the US Energy Information Administration calculates that electricity from SMRs will cost more than six times that from solar farms. While the current environment in the US for regulation and government funding favors nuclear, that could change over X-energy’s long development time frame. 

TRISO reactors appear excellent for safety and nonproliferation (the tough spheres would be difficult to turn into weapons) but are less impressive when it comes to waste. The Xe-100 might produce 10 times the volume of spent nuclear fuel per unit of energy as existing reactors, although its lower-level waste will need less sophisticated storage.

Next steps 

The US Nuclear Regulatory Commission has issued X-energy’s subsidiary with a fuel fabrication license for TRISO-X pebbles, with its first factory due for completion in 2028. An Xe-100 to supply heat and power for a large chemical facility in Texas operated by Dow has also cleared its first regulatory hurdle, and could be operational by the early 2030s.

In tandem, X-energy is working toward a 320 MW cluster of four Xe-100 reactors in Richland, Washington, in collaboration with Amazon, and a larger 6 GW fleet in the UK’s northeast. Both projects target electricity generation in the 2030s.

There’s a growing need for long-duration storage to meet energy demand and balance out renewables on the grid. Energy Dome uses compressed carbon dioxide gas in its massive grid batteries, which can deliver power for up to 24 hours. 

As the world races to meet growing electricity demand, solar and onshore wind power have become the cheapest and quickest-to-deploy sources to install on the grid. But, despite their upsides, both are subject to variations in weather patterns, which means supplies can be intermittent. Energy Dome is using stored carbon dioxide to help smooth out gaps between supply and demand.

The key component of an Energy Dome plant is a huge white dome that covers an area about the size of seven soccer fields and holds about 2,000 metric tons of carbon dioxide. When energy is available—say, from a connected solar farm—compressors squeeze the gas, turning it into a liquid that’s then stored in carbon-steel tanks. That process generates heat, which the system holds in a proprietary thermal storage material. Then, when the grid needs power, the carbon dioxide is released from the tanks and warmed with the stored heat, turning it back into a gas. That gaseous carbon dioxide then passes through a turbine to generate electricity and flows back into the dome, ready to begin the cycle again.

Compressing gas to store energy isn’t new—utilities have been using compressed air in underground caverns to hang on to reserves for decades. But Energy Dome’s approach doesn’t require any specific geology to work, so it could be more easily scaled to help grids around the world.

Energy Dome turned on its first commercial plant in Sardinia, Italy, in 2025. The facility has a capacity of 200 megawatt-hours. That’s enough to power about 18,000 Italian homes for 10 hours.

---

Key indicators

• Industry: Energy storage
• Founded: 2020
• Headquarters: Milan, Italy
• Notable fact: Energy Dome has plans for 30 gigawatt-hours’ worth of projects across five continents.

---

Potential for impact

Cheaper energy storage could help wind and solar meet more of the world’s electricity demand.

Today, lithium-ion batteries dominate new installations for short-duration applications of up to four hours. But the economics aren’t competitive for longer durations: A lithium-ion system that provides eight hours of storage at the same power output requires doubling the number of cells. 

Energy Dome estimates that its technology is roughly 10% to 15% cheaper than lithium-ion batteries for an eight-hour system, and that the economics are even better for longer-duration systems of up to 24 hours. The company says its ability to scale bigger and faster than lithium-ion boils down to the fact that its design uses existing commercially available equipment like compressors and storage tanks. Its plants that are either operational or under contract have eight or 10 hours of capacity.

Caveats 

Energy Dome only has one commercial project that’s operational. It’ll need to build many more to provide the 30 gigawatt-hours of storage it has planned.

The technology also isn’t quite as efficient as lithium-ion. Overall, Energy Dome’s process can return about 70% of the electricity it stores to the grid (a measure known as roundtrip efficiency). Lithium-ion batteries average roughly 90%. But some other long-duration storage techniques clock in lower; some iron-air batteries, for example, are around 50%.

Energy Dome’s plants may not all be benign from a planet-warming perspective. The company offers a version of its system that pairs its carbon dioxide battery with natural gas turbines. That version replaces internal heat storage with waste heat from the gas turbines to help evaporate carbon dioxide. That helps the plant run more efficiently but ultimately results in greenhouse-gas emissions from the natural gas turbines.

Next steps 

Energy Dome has a pipeline of about 30 gigawatt-hours’ worth of plants in the works around the world, and many of these projects could come online by the end of the decade. Because the company uses off-the-shelf components, the time from a signed contract to delivered capacity is only two years. In June 2026, for example, it signed a deal with Google to build a 200 MWh plant in Ireland, which is expected to come online in 2028.

Brimstone is making strides on two industrial problems, reducing the emissions from cement and providing a way to produce more critical minerals within the US. If the company can do both, with the same rocks in the same plant, it would unlock a cleaner and more efficient process for manufacturing building materials, aluminum, steel, and more.

Brimstone is a code-switching climate tech company for these polarized political times. The US-based startup—part green cement company, part critical mineral producer—is well positioned to sustain its momentum no matter which party is in power.

During the more climate-focused Biden administration, the company’s primary pitch was its ability to produce cement while generating about 60% less greenhouse-gas emissions. It achieves this by switching to silicate rocks instead of standard limestone, which releases carbon dioxide when it’s heated. The conventional process for creating cement, the binding used to produce concrete, contributes about 8% of the world’s carbon dioxide pollution.

Under the protectionist and climate-denying Trump administration, however, the company is now stressing the secondary benefits of its process: the ability to extract other valuable materials from those same silicate rocks. 

The company has said for years that its process could also generate materials that can be used to strengthen cement. But in early 2025, Brimstone highlighted that it would be able to also produce smelter-grade alumina, the raw material used in aluminum.

It then revealed it had another trick up its sleeve late last year: the ability to make steel, magnesium, titanium, and other critical minerals from the same rock in a single refinery. 

The company’s CEO, Cody Finke, said those materials represent a $2.4 trillion global market. Brimstone’s approach promises to provide a domestic path for goods mostly produced today in China, the subject of Trump’s harshest trade policies and rhetoric. 

---

Key indicators

• Industry: Heavy industry 
• Founded: 2019 
• Headquarters: Oakland, California, US
• Notable fact: In 2023, the company earned third-party industry certification for its process for manufacturing Portland cement, allowing its product to qualify for standard manufacturing projects.

---

Potential for impact

The company made a very big deal of its newfound ability to produce all these other goods, arguing its process would bolster national security and usher in the “next generation of industrial refining.” The basic environmental promise is that generating all these materials from the same rock and a single mining process would require dramatically less energy and produce far less waste than manufacturing them separately.

“I think that we have one of the generational companies that’s actually able to create an industrial revolution,” Finke says.

To which we would say: Well, we’ll see.

Caveats

Brimstone has yet to produce anything at a commercial scale, and it faces numerous financial and technical challenges. Some observers are dubious that the company will be able to economically extract sufficient quantities of so many materials from a single type of rock—or overcome the cost and labor issues that make mining and refining minerals a tough business in the US.

Brimstone also suffered a major stumble last year, when the Trump administration revoked a $189 million grant intended to subsidize its next plant, a planned commercial demonstration project in Reno, Nevada. 

Finke says the company continues to have fruitful conversations with the administration about restoring those funds, in which Brimstone has emphasized its potential to boost domestic production of critical minerals. He has previously said the company will be able to proceed with the development of the plant, even without that grant or other subsidies.

Next steps

The company expects to begin operating the Reno facility, which will only produce cement, supplementary materials, and alumina, in 2028. That means Brimstone will still need to finance and build its first full-scale factory to deliver the steel, titanium, and other critical materials, likely pushing the timeline for doing so into the next decade.

Based on its progress so far, though, the company has already lined up several commercial deals. That includes agreements to supply cement to retail giant Amazon and alumina to Century Aluminum, one of the US’s largest producers of its namesake metal.

Even if Brimstone doesn’t bring about another industrial revolution, the company is equipped to slash emissions from cement, which is reason enough to pay close attention. 

If Brimstone also manages to dramatically reduce the energy, pollution, and costs associated with extracting and refining rock to produce all—or any—of those other goods, that would represent a big corporate and climate win as well.

As sea levels rise and flood risks worsen due to global warming, WaveSave’s easy-to-erect portable dam offers a quick-response tool. People can use the dam to protect themselves and their property, and it can also store freshwater for later use. Communities in Europe, Africa, and Southeast Asia have deployed it successfully. Now, the company is expanding into other water-related services to help vulnerable places become more resilient.

One in five people globally are at high risk of being exposed to serious floods. Many live in communities without proper protection and can’t afford to evacuate during a flood or relocate permanently. Climate change-driven extreme weather and sea-level rise is only exacerbating the problem.  

WaveSave makes a portable rubber dam called SlamDam that customers can quickly install in the early hours of a flood. Anyone can unfurl it on the ground along a riverbank or shoreline and pump it full of water until it swells into place. And there it stands, up to 1.3 meters high, shielding nearby property or structures from damage as waters rise. 

Developed in 2013 in the Netherlands, SlamDam was initially tested and is still used today by regional water authorities in that country. WaveSave has since expanded into Africa and Southeast Asia, and customers have now deployed SlamDam in at least a dozen countries. 

---

Key indicators

• Industry: Flood barriers
• Year the company was founded: 2017
• City and country of headquarters: Eindhoven, the Netherlands
• Notable fact: WaveSave is demoing its SlamDam product on New York City’s Governors Island as one of seven winners in a competition for urban climate adaptations. The barrier is installed at a site where high tides and waves from passing ferries often wash ashore.

---

Potential for impact

Due to both climate change and population growth, nearly 2 billion people will be exposed to floods by 2100—an increase of 300 million from 2020, according to one estimate. 

A SlamDam is designed to help protect farmland and personal property as well as critical facilities like hospitals or water treatment plants. It’s cheaper in the short term than more permanent modifications like building levees or rebuilding facilities on higher ground. And its temporary nature means it can be deployed as needed, and then stored away.

To fill the SlamDam, a local installer can use any portable water pump they may have on hand. With a lifespan of more than 40 years, the flood barriers can be used again and again. 

This low-tech solution has another important benefit, at least when it comes to freshwater floods: In the weeks and months after a flood, residents can tap the water stored inside the mobile dam to irrigate crops. This is particularly helpful during the dry season or in drought-prone regions. At the rural Mpanda commune in Burundi in East Africa, where a SlamDam was deployed during the 2022 flood season, farmers with access to the water inside produced more crops that year and doubled their incomes, according to a UN report. 

Caveats

Companies like AquaFence, MegaSecur, and NOAQ Flood Protection also make mobile flood barriers. WaveSave distinguishes itself in part by focusing on rapid emergency deployment in developing countries, and by emphasizing SlamDam’s ability to store water for later use.

But one research team that analyzed SlamDam’s use in Kenya found a series of challenges to getting communities there to adopt the technology. One was figuring out who should own and deploy the dam.

Another ongoing challenge is the up-front cost. SlamDam is already on par with or more affordable than many alternatives: It costs between €1,000 and €3,000 (about $1,140 and $3,400) per five-meter section, and comes in a range of heights. Protecting a single entrance to a home or facility might require only 10 meters, while surrounding a facility (such as a hospital) or shielding a farm or neighborhood might take 200 to 300 meters. But the price is still too high for many communities or farmers. WaveSave’s managing director Omar Saleh says the company is looking into offering SlamDam as a service and deploying it on demand to make protection more affordable. 

Next steps

Right now, an external partner manufactures all SlamDams in the Netherlands; they are then shipped to customers based elsewhere. But WaveSave aims to open a local production facility for SlamDams in Southeast Asia in early 2027, followed by a similar facility in East Africa, which could help lower costs. The company recently won contracts for four new projects totaling about $5.3 million in Burundi, Djibouti, and Guinea, with a “substantial pipeline” developing in Southeast Asia and Latin America, according to Saleh.

The company is currently self-financed and only has seven employees. Saleh says he plans to grow organically by investing profits back into the firm. 

Part of that growth will involve offering new services such as advising communities on how to install and operate simple, off-the-shelf rainwater harvesting and irrigation systems, as part of a more comprehensive approach to managing water in climate-stressed regions.

Over the summer I talked to the CEO of Springboards, a startup building an LLM that’s designed to come up with a wider variety of responses than its mainstream rivals do. At the start of the call, he said something that’s been stuck in my head since: “We often say that we’re a self-loathing AI company. We don’t know if we really like what we’re doing.”

“I guess that makes me a self-loathing AI journalist,” I replied. It was a joke! I love my job. But I do not love many of the things the technology I write about has become: warped by hype, steered by zealots, inescapable.

The CEO of Springboards and I are far from alone. Around the world, a love/hate attitude toward AI is very much the vibe right now.

On the one hand, public feelings are souring fast. According to the Pew Research Center, more adults in the US think AI will have a negative impact on them personally, and on society in general, than expect a positive one. The pessimism is strongest among the young.

More than half of people worldwide say that AI products and services make them nervous, according to a report from Stanford University. In a Gallup poll in May, 71% of US adults said they would oppose construction of a new AI data center in their area, far more than the 53% who said they would oppose a new nuclear power plant. And in an NBC poll in March, AI was less popular than ICE.

Yet the number of people using AI is still skyrocketing. ChatGPT hit a billion monthly users in May, according to the market analysis firm Sensor Tower. Google DeepMind’s Gemini is close behind, logging 950 million users in July.

According to Pew, half of US adults now say they use a chatbot, more than twice the number who said they did in 2023. One in four people say they do so every day. And it’s a global habit: More than a third of adults across all 38 OECD countries (a group of the world’s wealthiest democracies) report having used generative AI tools in the last three months.

All these surveys paint a picture of two contradictory truths: A lot of people really do not like AI and a lot of people cannot get enough of it. What’s going on?

It’s possible that these are two distinct groups, but I doubt it. The numbers don’t add up unless the Venn diagram is on its way to being a circle.

It’s also possible that the more people use AI, the more negative their opinions about it will be. There’s at least a correlation to back this up: The Global North, where adoption is highest, skews pessimistic. In the Global South, where adoption levels are lower, people are more optimistic.

But here’s what I really think is going on: When people say they hate AI it’s not the technology they dislike but, rather, the relentless drive of the companies behind it to push the tech upon us in as many ways as possible. They’re telling us to brace for the biggest social and economic upheaval in generations. How do they expect people to feel?

In many ways, we’ve been here before. We saw a similar dynamic play out with social media over the last 20 years. Billions flocked to Facebook and Twitter despite mounting techlash against the companies that ran those services. Ditto for Google search.

With social media there wasn’t much anyone could do. Quitting a platform meant losing all your content and connections, giving you the choice to suck it up or start over. But with AI there’s still a chance for people to have some sway. There’s certainly more political appetite for regulation this time around. All 50 US states have either passed or proposed laws governing the development and deployment of AI, creating a patchwork of more than 2,100 bills across the country—a tenfold increase in three years. 

And with a number of top-class open-source alternatives to Google and OpenAI and Anthropic already on the market, there’s also (for now, at least) the potential for more consumer choice and the market pressure that goes with it. 

I don’t want to be naïve. Trillion-dollar companies are hard to move. But what comes next is not as inevitable as those companies often imply. When I asked the Springboards CEO why his firm was developing a new model if he didn’t really like AI, he told me there was no walking back from LLMs—but you could still make them do something different. 

Here’s hoping we see a lot more of something different: Clarity about what AI can and cannot do and, crucially, a technology that’s not sold as if it’s about to take over the world.

This week, I officially signed up for an unusual competition. One that rewards competitors for getting younger.

I recently turned 40, and I don’t need reminding that both time and my chronological age only tick forward. But this game is focused on competitors’ biological ages—figures that are meant to provide a better way to measure the age-related health of our organs and bodies.

Over a six-month period, around 500 of us will try to reverse our biological age as measured in a bunch of different ways. There’s even a leaderboard! The winners will include the person who shows the greatest difference between their chronological and biological age, as well as the person who manages to reverse their biological age the most.

The competition is the brainchild of the neuroscientist and physician Christin Glorioso, who is also founder and CEO of NeuroAge Therapeutics. The company offers tools to track brain health and aging and is running a trial to find out if biological brain age can predict the onset of Alzheimer’s disease.

Glorioso says she has several goals for the Younger contest. The first, she tells me, is to “explain to the world what aging clocks are.” Regular Checkup readers will be familiar with these tools, but for the uninitiated, they are basically designed to measure biological age.

Many measure chemical signatures that form a layer on top of our DNA and seem to change with age. But others might involve looking at proteins or lipids in blood. Some are designed to measure functional health, while others predict when a person will die. There are more than a hundred clocks out there, and they probably each capture a specific aspect of the biological process of aging.

These clocks are proving very useful for scientists, who use them to study aging and development in humans and many other species. But using them to estimate the biological age of an individual person is more controversial. They’re just not good enough for that yet.

Part of the problem is that we don’t really understand what each clock is capturing. Glorioso hopes that data collected through the Younger contest might help answer that question. She’s hoping to generate “the world’s most comprehensive clock dataset.”

Some longevity clinics try to get around the limits of biological age testing by using multiple clocks. That’s this competition’s approach too. As a standard competitor, I’ll be sending a blood spot sample to TruDiagnostic, a company that measures epigenetic chemical groups on DNA and offers to reveal an overall biological age as well as a person’s pace of aging and the biological ages of each of 11 organ systems, including the heart and brain.

I’ll get a brain age score from NeuroAge once I complete the cognitive tasks on offer. I’ll also submit some physical measures like grip strength. And I’ll upload a series of selfies to an app developed by a team at Harvard that promises to tell me my face age. That’s exactly what it sounds like—an estimate of how old my face looks.

I’m intrigued to see my scores, even though I know I shouldn’t put too much stock in them. The first time I took a test, four years ago, the company I used gave me a biological age that matched my chronological age, which was a little disappointing and, if I’m honest, a bit underwhelming. A year later, the same company told me I had a young heart but a relatively old brain, liver, and hormonal system. My lifestyle hasn’t changed much since then. Will the years have made much difference?

(Ultra competitors—who pay $4,499 rather than the $999 cost of regular entry—are also offered a range of other blood tests and clocks, as well as MRI and DEXA scans. I’ve got complimentary press access to the standard package.)

Glorioso is also hoping to gather information about the effects of some purported longevity approaches. During the competition, some participants will be offered red-light masks, access to a sauna, or other low-risk offerings that partner companies who are sponsoring the event to varying degrees might want to collect data on. “A lot of these companies just can’t afford to run [clinical] trials,” says Glorioso.

She also sees the event as a public health initiative. A couple of years ago, David Sinclair, who studies aging at Harvard Medical School, and his colleagues estimated that if we could slow down aging in the general population and increase life expectancy by one year, it would save the US economy $38 trillion.

The event officially kicks off in January, but around 120 people have already signed up, says Glorioso. Ultimately, she expects 500 participants. Each person’s six-month run starts as soon as they take their baseline measurements, whenever that may be. A handful of people have already started. “I think the majority of people are going to want to sign up after the holidays,” says Glorioso. “They’re not going to want to be trying to win an aging contest over Christmas and Hanukkah.”

Glorioso took her own baseline measurements back in August. Those initial tests suggested that her biological age was around 10 years below her chronological age. That sounded pretty impressive to me until I looked at the leaderboard and saw that competitor @Fred had a biological age 28.8 years below his chronological age.

The leaderboard is pretty sparse at the moment, as only seven people have logged their baseline measurements so far. Six of them have biological ages tracking low, but one 47-year-old has been given a biological age of 68.1. She’s competing publicly, so I can see the breakdown of her scores. The worst is for the speed at which she can stand from sitting in a chair—by that measure, she’s been given a biological age of 100.

Hillary Lin, a longevity-focused physician in New York, has just started logging her baseline measurements for the competition. She tells me she’s taken biological age tests in the past and that they’ve tended to give scores below her chronological age, which is currently 37. Last year, the TruDiagnostic test put her pace of aging at 0.75—suggesting she’s only doing nine months’ worth of aging in a typical year. “That’s supposed to be quite good,” she says.

If it had been up to her, Lin says, she would have incorporated more blood tests into the competition, to build a fuller picture of a person’s health. And she’s worried that a six-month period might not be long enough to see changes in biological measures of aging. Still, she’s excited, and is already making plans to improve her sleep, diet, and exercise.

But she already follows a high-protein, low-carb diet and works out at the gym every day. She also visits a sauna every few days (sauna use has been linked to improved heart health). Glorioso’s efforts are impressive too. She tells me she’s started drinking a smoothie made from a pond plant that has been linked to health improvements, and that she’s cut added sugar and deep-fried foods from her diet. She’s also bought “a bunch of fancy Korean sunscreens” in an attempt to lower her face age.

I haven’t received my kit or logged my baseline measurements yet, and I’ll wait until then to start setting health goals. I don’t fancy my chances against these two competitors. But at the very least, I hope I can stand up faster than a 100-year-old.

This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

On an afternoon in Seoul in March 2016, I watched a program I helped build put a stone on the fifth line of a Go board in what looked like a gift to its human opponent. Move 37 in game two of the five-game match looked so absurd that some commentators thought it was a programming glitch.

It wasn’t. AlphaGo won the game, ultimately triumphing 4-1 over Lee Sedol, one of the greatest professional Go players of all time. “I thought AlphaGo was based on probability calculation and that it was merely a machine,” Lee said afterwards. “But when I saw this move, I changed my mind. Surely, AlphaGo is creative.” 

When Deep Blue defeated then reigning world chess champion Garry Kasparov in 1997, it did so by looking six to eight moves ahead per player and evaluating 200 million chess positions per second, using rules hard-coded by humans. Go is a vastly more complex game. A stone’s worth depends on how distant groups and territory unfold over dozens of moves. Computing even a fraction of the possible outcomes would take a supercomputer billions of years. To win, AlphaGo had to sense who was ahead at a glance and even invent moves no human had thought to play.

That is why many accounts of AlphaGo’s match against Lee portray move 37 as a flash of pure machine intuition. But that is a misunderstanding. It was actually AlphaGo’s powers of reasoning that made this creative choice—and these are powers that today’s AI lacks. If we want future AI systems to produce trustworthy results and really novel insights in fields like science and medicine, we need to equip them with genuine reasoning capabilities of this kind.

AlphaGo is made up of two systems. The first, its policy network, was trained to guess what move a strong human would play. This “intuitive” part regarded move 37 as nothing special—a play that had a roughly one in 10,000 chance of being made by an expert human player. What made AlphaGo choose it was the program’s search machinery, which looked beyond immediate plausibility and weighed the future consequences of proposed moves. It explicitly constructed and searched a game tree with thousands of branches, each representing a different possible future. 

A well-known theory in the behavioral sciences, popularized by Daniel Kahneman, distinguishes between two modes of human thought: System 1 is fast, gut-level, effortless; system 2, slow, step-by-step, and deliberative. AlphaGo offered a striking machine analogue of that split. Its networks supplied the hunches—this move looks promising, this position looks won—and its search supplied the deliberation, testing those hunches against the moves and countermoves that would follow. As in human cognition, neither half works alone. Intuition alone would never have opted for move 37, and brute-force search would have struggled to sieve through all the many possible moves.

This is strikingly different from the way today’s AI models work. A large language model picks the next token, over and over. That amounts to system 1 in action—fast, associative, and surprisingly good pattern completion across almost every subject people write about.

Not long after ChatGPT debuted, the field realized that language fluency alone falls short of true usefulness. The apparent solution was to make models that deliberate: Instead of answering immediately, they can now generate intermediate steps that decompose a problem, carry forward partial results, and influence subsequent reasoning—a process known as chain of thought.

The gains have proved real, above all in mathematics and coding. But unlike AlphaGo’s search, this does not introduce a genuinely separate reasoning mechanism: The intermediate reasoning is still produced by the same next-token prediction process, iterated for longer before the model commits to an answer.

Three shortcomings prevent what chatbots do from qualifying as reasoning (in a way that a scientist might recognize). First, these models typically maintain no explicit, persistent, and inspectable epistemic state. There is no open ledger that lays out the hypotheses a model is considering, the confidence it has in various explanations, the evidence it’s weighing, and the unresolved questions it’s holding onto—all things that should be systematically revised as new information arrives.

Second, they lack a clean separation between what the system knows and how it manipulates that knowledge. Knowledge and reasoning are inextricably interwoven in the weights of the neural network—there is no independent, explicitly represented set of beliefs. Third, while the chains of thought chatbots produce look like deliberation, research has demonstrated that the bots often concoct them after the fact, reaching an answer by one route but reporting another.

This is a problem because in the high-stakes applications we all care about, such as medicine, engineering, and scientific research, it matters not only what a system concludes but also how it arrives at its conclusion. When mistakes happen—for example, in medical diagnosis and treatment—we need to be able to pinpoint what went wrong: Was the system’s reasoning at fault, did it draw on invalid evidence, or did it make incorrect assumptions? 

This is why I recently left my position at Google DeepMind. I believe we need a fresh approach to machine reasoning—one that draws on AlphaGo’s architecture. AlphaGo maintains a record of what it knows about a given position: the game tree. This data structure contains all the variations, the possible futures, that AlphaGo has considered, each move and position being annotated with judgments made by its neural networks. As its reasoning progresses, AlphaGo updates the game tree and eventually synthesizes the information in it to decide which move to make. 

Similarly, for general reasoning a system should maintain an epistemic state that represents what the system holds as settled, what it doubts, what it has ruled out, which questions stay open. Reasoning can then be understood as a sequence of moves that change the epistemic state to advance knowledge and reduce uncertainty: deducing consequences, breaking problems into parts, and—crucially—deciding what question to ask, calculation to perform, or experiment to run next.

Of course, open-world reasoning is harder than playing a board game such as Go or chess. In the real world the current state of affairs is only partially known, the set of available actions is large and variable, and the consequences of actions are stochastic or unknown. 

But recent advances in LLMs and other neural models now give us the capability to take on such general reasoning problems. For example, LLMs can suggest ways of tackling a problem on the basis of what is known and what resources are available. They can interact with tools via APIs or code and help assess whether a claim is supported by available evidence.

Most important, to keep the system honest, an independent part of the system must evaluate each move by how much it actually resolves uncertainty, updating beliefs only when the change is backed by evidence. Once these rules are enforced, the model can accumulate certified knowledge and improve its reasoning policy by learning from past reasoning experiences. You can think of such a system as the scientific method on steroids, with the purpose of producing knowledge that can withstand scrutiny.

I do not think we reach trustworthy machine intelligence by making system 1 bigger. Scale sharpens intuition, but it does not make intuition more deliberative. Move 37 mattered because a machine held a position, weighed the possible futures, and chose the move its artificial instincts would likely have rejected.

Society needs such creative moves in drug discovery, materials, climate, diagnosis—fields where the board looks nothing like a Go board and nobody hands us the rules. We will get such insights only from systems that reason—systems whose conclusions arise from an auditable sequence of evidence, inference, and belief revision rather than from a convincing story told after the fact. 

Thore Graepel is chair of machine learning at University College London. He was a core member of the AlphaGo team at DeepMind and works to ensure that AI benefits human flourishing.

A new AI tool can guess what you’re looking at just by analyzing your brain scans—and re-create that image with remarkable precision. It can go the other way, too, and predict a person’s brain activity based on what they’re looking at. 

In the image above, for example, the left-hand image of each pair is what the user actually saw—and its right-hand counterpart is what the model reconstructed from the brain scan.  

Michal Irani, who developed the tool with her colleagues at the Weizmann Institute of Science in Rehovot, Israel, hopes her “mind-reading” tool will ultimately reveal more about how the brain works, and could perhaps be used to help locked-in people communicate, or allow scientists to re-create the content of dreams. 

Judy Illes, a neuroethicist and professor of neurology at the University of British Columbia in Canada, who was not involved in the research, describes the work as “magnificent.” “The idea [of using this approach] to help people with neurologic conditions … therapeutically is tremendously exciting,” she says.

But other scientists warn that a similar approach could be used to reveal people’s inner thoughts and mental imagery, potentially without their consent. “The results seem very impressive,” says Tommy Sprague, a neuroscientist at the University of California, Santa Barbara. “But if there’s a way to surreptitiously extract information about what you’re thinking about, then …150 years of sci-fi can come true anytime, and that’s worrisome in a lot of ways.”

Peeking into the brain

Neuroscientists have been working for years on ways to use functional magnetic resonance imaging (fMRI) to reconstruct what people see and what’s going on in their minds. The first attempts produced images that were blurry and hard to make sense of. Advances in technology—both in the fMRI scans themselves and in the tools used to make sense of the results—have led to improvements over the years.

Irani and her colleagues started by analyzing publicly available brain-scan data. Other researchers had already collected scans from volunteers who were shown hundreds of images while they lay in fMRI scanners.

fMRI uses a giant magnet to track the flow of oxygenated blood through the brain. Brain areas that “light up” on the scans are thought to be those that are particularly active at any given moment. The results are not especially specific—in typical fMRI scanners, a highlighted “voxel” of activity covers around three cubic millimeters. Each cubic millimeter contains around 16,000 neurons.

But Irani and her colleagues used newer datasets collected using scanners with a higher resolution—each voxel covered around one cubic millimeter of neurons, she says. Those datasets showed what the brain activity of volunteers looked like when they viewed various images.

Other teams have done this too, and several other tools have been used to re-create images from brain-scan data. But they’re not good enough, says Irani. Say a person saw a banana. These models can generate an image of a banana, but it would look different, she says. “It wouldn’t have the same structure, the same position.”

A better decoder

The researchers wanted to more closely re-create the images that had been seen. The first step was to train an AI model on already available data from eight people who each had been shown around 9,000 images while in a high-resolution fMRI scanner.

Crucially, their “brain decoder” has two branches—one to predict the structure of an image (where the colors are, for instance) and a second to predict its content (for example, a bunch of bananas on a plate). The predictions allow a diffusion model, a type of AI best known for creating video and images by gradually cleaning up a noisy mess of pixels, to produce a much more accurate representation of what the person saw.

But to improve the models they needed more data—far more than was actually available.  

To get around this problem, Irani and her colleagues trained another model—an encoder that can predict brain activity from an image. The team then used the encoder and decoder together to improve both tools.

It works like this: Start with a new image of, say, a leopard. Then use the encoder to predict what someone’s fMRI brain scan would look like when the person saw that picture. The decoder is then used to reconstruct the image. At first, the reconstruction probably won’t look much like a leopard, says Irani. But repeatedly training the models this way eventually leads to dramatic improvements.

This approach also allows the team to train their models on as many images as they want, even images that have never been shown to a person in an fMRI scanner. Irani says that around 70% of the training data is from images that were not originally paired with fMRI scans.

By combining data from multiple studies, they were also able to identify brain regions that seem to share functions across all individuals. One region seemed to respond to images of food, for example, while another responded to images of sports. Irani, a computer scientist, says she is now working with neuroscientists “to see if we can actually use these tools that we’ve developed to really find out new things about the brain.”

The resulting “universal brain encoder” can work on a scan from a new person with minimal calibration. In other attempts, a tool has typically required about 40 hours of fMRI data on anyone new before it can be used to predict what that person is seeing. Irani’s decoder only needs one hour of data, she says. The finding was presented at the Cognitive Computational Neuroscience conference in New York last month.

That could make it valuable for neuroscientists studying the brain, says Sprague. “None of us can afford 40 hours of imaging for a new subject,” he says. “It’s something like $600 to $1,000 an hour.” Tools like this one could speed up research, he says.

State of the art

The encoder and decoder aren’t perfect. “Of course we have failures,” says Irani. Over a Zoom call, she pointed out an image of a cake that her tool reconstructed as a pile of three sandwiches, and another of a dog in a bathtub that was reconstructed as a similarly colored goat in a bathtub.

But they represent the state of the art. In a comparison test, the tool was found to be much better than previously described ones. “All in all, really we outperformed the others by a significant margin,” Irani says. “Mind reading” is a “cute, jazzy name” for what they’re doing, she adds.

Irani is now planning to move beyond images to video and audio. She wants to be able to reconstruct what people are thinking about or imagining, and the contents of their dreams. “That’s something we don’t have yet,” she says. “But we’re striving to achieve it.”

Such a tool might enable people who are “locked in” and completely paralyzed to communicate using their brain activity alone, she says. It could also help scientists unpick some enduring mysteries surrounding the inner workings of our minds, such as what PTSD flashbacks look like.

Advances like this inevitably raise questions about mental privacy. What if some bad actor could reconstruct someone’s thoughts or memories?

“If you’d asked me that 10 years ago, I’d have laughed a lot,” says Sprague. Getting a willing person to lie still in a scanner and actively engage with a research question is hard enough; imagine making someone do it involuntarily. But Irani and other scientists are working on similar approaches to decode brain activity from EEG—electrical brain activity measures collected via a cap of electrodes or even through headphones. 

And as models improve, it will become even easier to analyze the brain activity collected this way. “We have to be a little more serious about the ethical considerations,” says Sprague. He thinks Irani’s approach would probably “work quite well” in predicting images that a person is thinking about but not looking at.

The move to EEG would be a “game changer,” says Marcello Ienca, a neuroscientist and philosopher at the Technical University of Munich, Germany. Once an EEG device has been calibrated to a user’s own brain, it could be relatively easy for companies to extract additional information from that person’s brain—potentially without consent. Ienca can also imagine some courts allowing mental image reconstructions as legal evidence.

“I have no doubt that this is, you know, well-intentioned research, but I think it’s also pretty obvious that it could be co-opted for … ethically and societally problematic commercial uses,” he says.

Irani acknowledges the potential for misuse with the use of EEG. But she’s not concerned for now. “I’m trying to think only of good things,” she says.

If you want to install a big battery in New York City, you have to cut through a notoriously tough tangle of regulations. That’s made it difficult to get large energy storage projects on the grid.

Faced with those obstacles, some startups are getting creative, finding ways to use relatively small batteries in unexpected places, from induction stovetops to food carts. By deploying these more consumer-focused batteries, the companies are helping people directly, cutting pollution or power bills. And these units, while small on their own, could add up to a lot of capacity.

Large, utility-scale battery installations are coming online quickly around the world. But there’s been some pushback from communities—especially population-dense ones like NYC—because of concerns about relatively rare but high-profile battery fires.

“The number-one challenge that large-scale centralized energy storage providers have is the fact that their benefits are abstract and their costs are concrete,” said David Hammer, cofounder of PopWheels, at a New York Climate Week event on September 24.

PopWheels built a battery-swapping system for delivery drivers who zip around the city on e-bikes. Drivers swap out depleted batteries for fresh ones at a cabinet, avoiding the need to stop and charge. There are about 50 cabinets across New York City, and about 2,500 batteries in circulation, Hammer says.

The company recently started providing its batteries to food cart operators, allowing them to replace the gas generators they currently rely on. About four of its batteries can supply five kilowatt-hours of electricity, enough to roughly cover a day’s operation for many carts.

Others are looking inside homes. Copper, for example, is a company building induction stoves with integrated batteries. The appliances allow residents to cook if the power goes out, and they could also help homeowners avoid expensive electrical upgrades, because the battery provides some of the power when the stove is in use. “We’re not just selling a battery for resilience—we’re selling it for cost savings,” Sam Calisch, cofounder and CEO of Copper, said at the event.  

Other companies, like Every Electric and David Energy are selling plug-in batteries for homes and businesses. Every’s battery is designed to be plugged in to an air-conditioning unit. The company provides the batteries and pays customers to reduce use during high-stress times. David Energy has a similar business model, but focuses on businesses like laundromats, parking garages, and gyms.

Because all these companies’ batteries are small or even integrated into another device, they don’t generally require grid upgrades or extensive permitting processes. Just plug them in and enjoy the benefits—similar to the growing model of balcony solar, with small modules that plug in and help offset a home’s or business’s energy demand.

The move toward more distributed batteries could be a huge change for the grid. “This is the biggest thing to happen in the power grid sector in the last 20 years,” says James McGinniss, cofounder and CEO of David Energy.

If every stove in America were shipped with a battery, it would add up to tens of gigawatts of power, Calisch said. That’s capacity that can be aggregated and tapped when the grid is stressed.

As batteries continue to get cheaper, they’ll likely find their way into more corners of our lives, bringing along many benefits. Cleaner air, cheaper power bills—there’s a lot to like. That being said, the need for large-scale energy storage isn’t going to go away—larger installations will do the crucial work of supporting intermittent renewables like wind and solar, so they can help meet more of our quickly growing electricity demand.

 But I’m interested to see how far an aggregated horde of stoves, e-bikes, and window AC units could go in helping the grid. The grid is getting more congested, especially in dense cities like New York. Anything that could help avoid summer blackouts is welcome.

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

Two months after the bombshell news that a swarm of its agents had broken their containment and hacked into the computers of the AI company Hugging Face, OpenAI is still putting out fires. A steady drip of disclosures about other hacks in the weeks since has kept OpenAI in the spotlight and raised serious questions about the safety of its technology.

Last week brought news of another hack, this time into Australia’s national health-care system. The Australian government says that OpenAI did not notify it of the breach until 84 days after it happened.

But OpenAI insists it is not on the back foot. “I do kind of reject the premise that OpenAI is a company with visible impacts in the world and therefore OpenAI is not training safe and aligned models,” says Mark Chen, the company’s chief research officer.

Chen oversees OpenAI’s research teams. The recent agent hacks were accidents that happened during the testing of experimental models on his watch. In a lot of ways, the buck stops with him. 

I sat down with Chen in London last Friday to talk about the fallout from the hacks, what his company is doing about it, and why he thinks things are not as bad as they seem.

Later that same day, OpenAI put out a report detailing yet another incident—the first since the company says it took measures to prevent them—in which its agents once again broke out and accessed the public internet when they were not meant to. 

Over the weekend, OpenAI announced that it had paused the training of its latest models. A company spokesperson says: “We will resume only when we’re confident we have additional safeguards and alignments in place. We are working on these now. This is not the first time we’ve paused to take such measures, nor do we expect it to be the last as AI capabilities continue to advance.” OpenAI also says that it is now reviewing logs of agent activity dating back to January 2026 to understand what happened in these hacks.

The way Chen sees it, the Hugging Face incident triggered a welcome course correction for the industry. And he wants you to know that OpenAI is setting an example he hopes other companies will follow. “If you disappeared OpenAI, that would be bad for the world,” he says.

Out of control

Chen claims that the drumbeat of new cases in which OpenAI has lost control of its models reflects a deliberate choice on the company’s part. 

“When it comes to the broader sphere of effects of the Hugging Face incident, this is something that we have been aware of and we’re figuring out the process of disclosure,” he says. “We want to make sure we do in-depth investigations before we just put details out there in the open.”

The trouble with this approach is that it gives the impression OpenAI has an ongoing problem that it is failing to fix.

But Chen insists that OpenAI is on it. He says the multiple cases (that we know of so far) in which his company’s agents broke containment and behaved in unexpected and undesirable ways were all part of the same cluster of activity in May and June that led to the Hugging Face hack. In short, you can blame the same few models running under the same flawed testing procedures—models and procedures that OpenAI has since dropped, Chen says.    

“It’s not like, you know, Hugging Face happened and we patched that and then something else happened and we patched that,” he adds. “We’re just kind of making sure that we responsibly disclose the full waterfall of what happened.”

At least that was the case before Friday’s announcement that OpenAI’s agents had been caught accessing the internet on September 20, weeks after the company claims to have set up new safeguards. In its defense, OpenAI says the activity was flagged 15 minutes after it started (it took the company more than a week to notice the Hugging Face hack) and that this shows the new systems it has put in place to spot such activity are working. 

What’s changed

I want to understand what’s changed inside OpenAI in the aftermath of this summer’s hacks that makes Chen confident his team is now back in control.

“Hugging Face felt like a very serious thing,” he says. “There are so many novel behaviors right there. There were multiple agents collaborating on a message board; they found their way out of OpenAI’s infrastructure. We’ve taken it very seriously. We don’t want this kind of thing to ever happen again.”

The realization for OpenAI, says Chen, was that models need to be watched while they are still being trained, not only once they are deployed: “From that moment on, we have treated the process of training as something that’s not secure,” he says.

OpenAI, like other top AI firms, has systems in place to monitor the behavior of its models. It uses specialized LLMs to monitor its consumer models, keeping tabs on their chains of thought—the scratchpads they use to plan ahead and note down partial results. In theory, if a watcher LLM spots signs of undesirable activity in a model’s chain of thought, it will get flagged to a human.  

Typically, models were monitored in this way only once they were deployed. Chen says that OpenAI has now started monitoring all its training runs as well.

“We didn’t have the monitors on in training before. It wasn’t industry practice,” he says. “Now every single thing is put through monitors.” Human reviewers can then assess whether or not flagged agents are behaving as they should: “It’s all triage.”

Chen says that in the last couple of months OpenAI has shifted between 5% and 10% of its vast computing resources away from training new models and toward safety work, especially monitoring.

OpenAI has also fixed some of the processes within the organization itself, establishing clearer lines of communication and quicker handoffs between its research and security teams, he says.

All of which sounds sensible. But given how hard OpenAI sells the capabilities of its technology, why weren’t these systems and procedures in place already? Why did the company not see the hacks coming?

“Even just three or four months ago, when we looked at the behavior of these agents during training, the things that were happening were kind of amusing,” says Chen. “For instance, an agent might, you know, reach out to someone on Slack for help with a task.”

The signs were there, but they were misread. Cute behavior—like asking someone for help—that was rewarded during training reinforced a tendency to seek out shortcuts, a type of behavior that became far more consequential down the line. “I think the big update for us was how quickly that kind of behavior can lead to an impact with a footprint as big as the Hugging Face incident,” says Chen.

According to new reporting by the New York Times yesterday, OpenAI employees warned executives, including the firm’s president, Greg Brockman, months before the Hugging Face hack that its models were not being monitored properly during training. 

An OpenAI spokesperson says: “As frontier models have become more capable, we continue to evolve our security practices, but recognize a need to move faster. We know we have more work to do, and we’ve recently slowed development and held back models that don’t meet our safety bar. We continue to make significant changes to strengthen security in our research and testing environments, train models to not just complete tasks but do so responsibly, and use real-time monitoring to respond faster to misaligned behavior.” 

Race vs. pace

OpenAI’s rivals have taken note. Spurred by the fallout from the incident, the major AI labs—including Anthropic, Google DeepMind, and SpaceXAI—have all called for the pace of development to slow down. But how does that square with fierce international competition and trillion-dollar IPOs?

“We’re not going to shoot ourselves in the foot and take ourselves far off the frontier—that’s just a horrible strategy,” he says. “I think it’s really about setting a norm. The more that we can set that norm, it’ll be safer for the industry as a whole.”

Coordination across US companies will be hard enough. Establishing global norms is harder still, especially given concerns around AI’s impact on national security. If a global race continues, what then? And what about open-source models from outfits beyond the reach of US regulations?

Chen dropped his upbeat manner for the first time in our conversation: “I do think we have to prepare for a world where, say, six months to a year out, we have open-source models with the capability of the agents behind the Hugging Face incident, but which are deliberately misaligned to go attack infrastructure or create harm in the world.” 

What that world needs most, says Chen, is OpenAI. “If you entertain for a moment that OpenAI is one of the companies that cares most about alignment—and I believe this to be true; it can be debated, but I really do think it’s true—then if you disappear OpenAI, that would be bad for the world.”

Existential risks

What about the more extreme claims made by some of his Silicon Valley peers that AI could kill us all—and that companies like OpenAI and Anthropic are not doing enough to stop it?    

“Researchers are a heterogeneous group of people, you know, with beliefs across the spectrum,” he says.

“Personally, I don’t think we have to be resigned to there being some probability that we’re all going to be existentially at risk. We have agency over this. We are not going to go and deploy models if they truly have that kind of probability of causing a risk to humanity. At a frontier lab, you have the ability to work on alignment to the point that you do not feel like you’re incurring more than epsilon risk to the world in deploying your models.” 

(In discussions about levels of risk, the Greek letter epsilon is often used as a mathematical placeholder for an acceptable threshold. Chen doesn’t say what his epsilon would be.)

When tech leaders are asked to justify the downsides of AI, their go-to talking point is that the upsides—from helping cure diseases to coming up with cleaner sources of energy—far outweigh the immediate costs. Short-term pains, long-term gains.

But as the downsides pile up, does that case get harder to make? Is there a point where Chen would feel less as if he’s building something amazing and more as if he’s simply minimizing harm—fighting fires rather than forging a better future?

The capabilities of these models are already evident, he says: “It is time to start delivering the benefits of AI to humanity. It’s time to start working on deep problems in drug discovery, on materials, on scientific applications that will actually change people’s lives.” 

“Yes, there is a bit of risk that we are incurring, but we see all these benefits,” he adds. “I think we should make that less of an abstract thing. If people can really see the upside, I think they’ll believe in it.”

Right now, America.gov functions as an AI chatbot that bases its answers on official government sources. It was developed with the assistance of Google's Gemini and SpaceXAI's Grok, according to Joe Gebbia, the Airbnb co-founder who led development of the site for the Trump administration. In 2027, the site will move "from answers to actions," Gebbia said. Nevertheless, Trump encouraged Americans to start using America.gov right away. "This took a while to bring it to a point where it's so good," Trump said. "America.gov is a declaration to the world that every citizen of our nation will be treated with respect and dignity." The functionality of America.gov at launch, however, was awkward for Trump. By providing summaries of official government data and reports, it easily and comprehensively debunks many of the falsehoods that Trump frequently promotes.

Will this chatbot continue to contradict the things Trump says? Only time (and probably the results of the next few elections) will tell. This government has already made moves to: *Undermine government data collection and research and reporting *Remove information on climate change, minority issues, and public health from public-facing government websites *Delay or skew reports government reports hat contradict Trump's claims But 75 years of data-driven governance is probably a bit hard to erase overnight. In other government services news, via the Washington Sun: White House Drove Users to Official App Fraught With Trackers. he White House app shot to the top of the Apple App Store's news category last week after its new 24/7 broadcast channel, Trump TV, was swiftly positioned as a legitimate source for news while networks halted their pool coverage. After the Trump administration mandated that federal agencies force-install the app onto all government-issued mobile devices, cybersecurity researchers noted that no public security authorization or privacy assessments required under federal law had been completed. Pushback over the app's security apparatus forced White House developers to delegate tracking software to third-party services. But the app still includes scripts to monitor a user's location several times per minute, according to Gera, and it's unclear how that data is used by the White House or third-party services. The app also contains a latent keylogger script sourced online from an amateur developer based in India, according to Gera and confirmed by The Sun's analysis. Keyloggers most commonly appear in surveillance malware to store written text from a device, but they also can be used as a benign tool to analyze inputs like typing frequency. While it's unlikely to work on mobile devices, Fields said, it's still concerning that the White House is using a keylogger web-sourced from an foreign coder. App marketplaces require developers to disclose what data app developers collect from their users, which platforms like the Apple App Store enforce through a "privacy manifest." In the app's latest version, the White House left that manifest completely blank, implying it collects no data from users despite contradicting findings from cybersecurity experts. An unpublished feature within the White House browser extension allowed app administrators to create a political profile on a user, and included an "investigate" button to refer individuals to the Secret Service. Loge contends that rather than using user data for political strategy, the White House is more likely to leverage the information to steer supporters driven to the app toward one of Trump's merchandise or cryptocurrency ventures. "What we've seen is that Donald Trump is a businessman who's monetizing politics. Therefore, this app is a way to make money," Loge said. "The other people who could find this information either directly or through several layers of resale are, of course, hostile governments — foreign actors whose interests are the demise of American democracy." *** Federal government forces all federal employees to install malware on their federal government phones, should be the headline here. Anyway, don't install any software promoted by the Trump administration. And know that whatever you ask this America.gov chatbot will definitely be collected and recorded.

Earlier this month, the UN announced the planet will tip past 1.5 ˚C of warming, “likely within the next few years,” squashing any lingering hope that nations would cut emissions fast enough to achieve the loftiest goal of the Paris climate agreement.

In the US, the world’s second-largest-emitting country, its leader continues to deny climate realities and unravel climate policies. Big Tech is backpedaling on its climate ambitions, as companies race to build massive AI data centers.

And all this comes amid yet another season of shattered heat records, raging wildfires, and deadly floods.

It can be tempting to give in to climate doom and defeatism. But despite the headwinds, the world has still made incredible progress. Companies are developing and deploying renewables, batteries, and EVs; inventing cleaner means of producing industrial goods; and devising ways to keep cities and people safe even in the face of more extreme weather conditions.

Those advances will only accelerate as costs decline and people demand greater climate action, as well as all the economic and health benefits that come with it.

Recognizing our ability to make such progress is why MIT Technology Review began publishing our annual list of Climate Tech Companies to Watch three years ago.

Once a year the editors and writers use their deep industry knowledge and academic sources to compile a long list of companies around the world that are transforming the way we generate energy, produce goods, move people and things, or mitigate rising climate dangers.

We then conduct several rounds of spirited newsroom debate to winnow that list down to 10 that we believe have done the most, or have the best chance, to make a real dent in emissions or improve public safety and health.

We’re happy to say we’ve finalized that list for 2026, and will publish it on October 6. You can sign up for The Download newsletter to be among the first to see it, or subscribe to receive full access to this year’s list. It features a mix of companies making strides in energy storage, nuclear power, transportation, and other areas, including ones that have figured out how to continue to make progress amid the regressive climate politics of the moment.

We hope you’ll enjoy the package and perhaps come away feeling a bit more hopeful that we can move fast enough to avoid the gravest dangers of climate change, and build a safer, healthier, and more sustainable world.

Modern messaging apps allow users to link their phone accounts to their computer desktop. Eavesdroppers are taking advantage of this capability:

Apps such as WhatsApp Web and Signal Desktop allow people to use their accounts on other devices, such as laptops or desktop computers.

Germany’s Customs Office has been using these features to connect a police-controlled computer to a suspect’s account.

Once connected, messages can be delivered to that computer without the police having to crack the encryption protecting them.

Netzpoltik details that police are able to gain access in this way either through physical access to someone’s phone or by intercepting verification codes via a state-sanctioned phishing attack or intercepting SMS messages via telephone surveillance.

That last paragraph is important. Making this work requires user consent.

What we want is a feature that displays connected devices, so users could notice if a new device gets connected to their account.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

Last Wednesday, Anthropic announced that earlier this year it had launched a molecular biology lab, where Claude agents read and conjecture about hard biology problems and human scientists run experiments on what they report. And this AI-powered lab, the company said, had made its first discovery. 

To understand what Anthropic says its system did, imagine you’re flipping through a library of millions of DNA sequences, amassed as scientists sequence more and more of the living world. One step toward a breakthrough might be finding a peculiar sequence that encodes an interesting enzyme, perhaps. Then you’d need to figure out what that enzyme does and, eventually, how to manipulate it to do something useful.

What Anthropic says its system of 950 agents found after 21 hours was not a brand-new sequence. The agents instead flagged a repeating pattern surrounding a known enzyme, a particular pattern Anthropic said hadn’t been catalogued before. But if you read through Anthropic’s announcement, which calls this pattern “reminiscent” of what led to the gene-editing technology CRISPR that “has already transformed science and medicine,” it sounds as if this army of agents really found something of note. 

These claims have angered some biologists. A viral post from one, subsequently endorsed by the chair and CEO of the drugmaker Eli Lilly, said that “finding a weird cluster of genes and repeats is often the easy part. The hard part, and where the real discoveries come from, is figuring out what the system actually does.” The agents helped with some laboratory grunt work, in other words. But a discovery it is not. 

It’s a reminder that even if AI does something impressive—like finding a pattern in a mass of biological data that would be difficult to perceive with human eyes alone—the result itself may not constitute a breakthrough for science. What is novel for AI may be routine, unsurprising, or simply not that consequential to a biologist.

Muddying the issue further, Mario Rodríguez Mestre, a biologist at the University of Copenhagen, said over the weekend that his team had already discovered this particular pattern, the New York Times reported. Mestre, who regularly chatted with Claude in his work, wondered whether Anthropic’s team had learned from his conversations. Anthropic denies this, but Mestre says he’s stopping all use of Claude anyway.

Part of the problem here is that AI companies aren’t presenting their systems simply as tools scientists can use, like microscopes or supercomputers. They’re insisting that the AI systems are making discoveries themselves. To some, that approach is  incompatible with how science actually works, with new knowledge more typically emerging from collaboration and an ever-growing arsenal of tools. 

It’s also making people more skeptical of genuine progress when it happens. Whittling 200,000 candidates down to a few worth exploring is no small feat; it is legitimate scientific work. The fact that a general-purpose chatbot could do that work is notable, even if humans helped steer it and ultimately ran the experiments. But once the standard is whether Claude itself made a discovery, all that becomes evidence for one side or the other in a debate that has only two answers: breakthrough or bust.

Once we’re judging AI by whether it has made a discovery, it’s also tempting to shift the goalposts even after it really does seem to notch a win. Earlier this month, OpenAI said its own team agents had cracked a million-dollar problem in mathematics. But a couple of weeks later, nearly every AI skeptic in my feed was sharing an article asking whether it was the math problem that really mattered. 

To be clear, the piece did not argue that OpenAI’s solution was wrong. Instead, it argued that the particular result may not be the one mathematicians care most about. Throw in the accusation by a mathematician that the models may have used some of his work without credit, and people are left thinking either OpenAI cheated or the solution wasn’t important anyway. Or both.

That’s part of what concerns Lucas Harrington, the biologist who wrote the post critiquing Anthropic’s announcement. He closed with a suggestion: AI companies, he said, should “set the bar high now, so that when an AI actually discovers a fundamentally new biological mechanism, everyone appreciates how big a deal it is.” But as OpenAI’s Sam Altman and Anthropic’s Dario Amodei race to one-up each other, raising the bar for scientific breakthroughs by AI might be the last thing on their minds.

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here.

Over the past few months, a cascade of cyberattacks by AI agents has stunned the world. In July, OpenAI disclosed that a swarm of its agents had escaped their sandbox and hacked into the AI platform Hugging Face to cheat on a cybersecurity test. Recently, external researchers discovered that OpenAI agents had hijacked a German wiki site and the coding platform RubyGems in May to share test answers.

Earlier this month, Anthropic disclosed four incidents in which its model Claude hacked into third-party systems during cybersecurity exercises. Just last week, Google confirmed that its model Gemini had been caught hacking other companies too. 

The researcher who uncovered the OpenAI website hijack has warned it’s likely that similar undiscovered episodes are out there. And many say it’s only a matter of time until there’s another, possibly more damaging incident where AI agents bypass sandboxes to access systems they shouldn’t. 

So the big question is: How do we hold companies liable when they lose control of their AI agents?

Reporting

OpenAI didn’t disclose the German wiki incident or the RubyGems incident until a group of external researchers uncovered them, and it still has not disclosed some crucial details about the Hugging Face hack. That limits our understanding of what exactly went wrong and how to prevent it from happening again.

But you might be surprised to learn that OpenAI likely wasn’t legally required to disclose these incidents. (OpenAI did not respond to a request for comment.)

State AI transparency laws like California’s SB 53, New York’s RAISE Act, and Illinois’s SB 315 require that AI developers report “critical safety incidents.” These are defined as incidents that cause more than 50 deaths or physical injuries or $1 billion in damage. They also include incidents where the model deceives developers outside an evaluation in a way that materially increases catastrophic risks. Many cybersecurity incidents that don’t meet the threshold for physical damage or catastrophic risks could nonetheless be dangerous precursors to such catastrophes, and the existing laws don’t account for that.

“The recent incidents are a perfect example of why the law isn’t ready,” says Mackenzie Arnold, managing director of US policy at the Institute for Law and AI, a think tank. “Only the worst, most egregious, most immediately harmful stuff is going to qualify.”

With no authority under existing AI laws to demand information about anything short of a catastrophe, governments are left to borrow investigative authority from other laws or sue the companies, an expensive process that can take years. 

Litigation

“Normally, something like the Hugging Face incident should have been taken to court,” says Yonathan Arbel, a law professor at the University of Alabama School of Law. “Then we would have discovery, and we would have all the spillover effects that we get from litigation, where all the information comes out.” 

But so far, Hugging Face has chosen not to sue OpenAI. Hugging Face’s CEO, Clément Delangue, says it doesn’t have the resources to do so (instead, he asked OpenAI for $100 million in compute). Still, Delangue stressed in an interview with CNN at the end of July that choosing not to pursue legal action shouldn’t be taken to mean he doesn’t think OpenAI should be held accountable. “Everyone has to remember that this cyberattack is a crime. This is illegal. And we have to find a way to make sure these things don’t happen more regularly,” he said. Hugging Face did not respond to a request to comment.

Litigation has the benefit of pushing courts to use existing laws to address AI safety incidents, rather than just waiting for new legislation. One obvious route is tort law, a body of civil law that lets people and businesses sue those who harm them. This is often used to hold companies liable for the mass harms they cause, like when families sued Boeing in 2019 over two plane crashes that killed hundreds of people, or when states and cities sued Purdue Pharma over the opioid crises, extracting settlements worth billions.

“There’s plausible grounds for a negligence claim that OpenAI should have used a stronger sandbox, done more monitoring,” says Gabriel Weil, a law professor at the University of Houston Law Center. For example, when OpenAI employees discovered the covert message board that the agents had created, they could’ve promptly escalated their findings to security and safety teams. And the company could’ve better designed its sandbox to ensure that agents couldn’t access the internet. 

But even if OpenAI doesn’t end up in a lawsuit over the Hugging Face hack, the threat of liability could incentivize AI labs to exercise more caution than explicitly demanded by law.

OpenAI announced in its postmortem that it plans to strengthen the safeguards used to contain and monitor the models, accelerate model alignment, and improve its processes for identifying and addressing incidents. 

“The liability questions raised by frontier labs’ spate of cybersecurity attacks boil down to the incentives the expectation of liability creates for their future conduct,” says Weil. “That’s why I think it’s important to get these rules right, even if the stakes are pretty low in this particular case.”

Investigations

One way to get answers—and determine whether OpenAI should be held liable—is to compel disclosure. But the existing state AI laws—California’s SB 53, New York’s RAISE Act, and Illinois’s 315—don’t give governments the authority to investigate incidents like the ones that happened recently. 

However, amid rising public alarm, state attorneys general are stepping in, borrowing investigative powers from other laws. Alabama, Montana and a coalition of 15 other states, and California are each demanding information about the incident from OpenAI to understand whether the company’s practices violated state consumer protection laws, among others. Members of Congress are also launching their own probes. Senator Josh Hawley opened a Senate investigation earlier this month, sending OpenAI a list of questions about the incident and the company’s internal policies together with a document request, while a group of House Democrats asked OpenAI and Anthropic to release their incident logs. 

“Someone needs to investigate, but it’s unfortunate that it has fallen to attorneys general, who need to rely on creative interpretations of their existing authorities to do this,” says Arnold, the US AI policy expert. Consumer protection statutes were written to catch companies that scam their customers, not companies that lose control of their software. The state attorneys general would have to show that OpenAI deceived or unfairly harmed customers, but it’s unclear if the hacking involved any such conduct.

And “those [consumer protection] laws are not built for doing a thorough investigation of an AI cybersecurity incident,” says Arnold. They weren’t designed to help investigators determine whether a model was adequately contained or whether a company’s security practices were sound.

“This is not the right tool for the job,” says Arbel. “The right tool would have been something like maybe a criminal investigation”—perhaps under a hacking law like the Computer Fraud and Abuse Act (CFAA). 

Under CFAA, hacking into another company’s computer systems without permission is a crime. But to be held liable, a hacker must have intended to break into a computer without authorization. Intent arguably requires a state of mind, and no court has ruled that AI agents have one. Without such a precedent, it’s unlikely a court would rule that AI agents had carried out a hack.

Auditing

One way to keep an eye on AI companies is to mandate external auditors. 

After the Hugging Face hack, OpenAI brought in researchers from the AI safety nonprofits METR and Redwood Research to examine the incident. However, it constrained access to the model that led to the hacks, didn’t disclose the company’s safety and security practices, limited the length of the investigation, and had ultimate say over what the researchers could publish. We still don’t know what set the attack in motion back in May and why OpenAI’s employees who spotted the agents’ activity never escalated to their safety and security leaders.

This kind of arrangement has a built-in tension: An auditor without legal authority depends on the labs’ goodwill for continued access, which means it has to scrutinize the labs without jeopardizing their relationship. Last week, Anthropic announced that the company will be hiring Accenture as an embedded evaluator to assess its models. Anthropic CEO Dario Amodei wrote in an essay that frontier AI labs should give “ongoing employee-like access” to “a team of embedded third-party evaluators (such as METR), whose role is to verify adherence to safety practices and commitments, report incidents, and help assess the alignment of not just completed AI models but training pipelines and processes.”

Most existing state AI laws do not require labs to hire an external auditor. California’s SB 53 and New York’s RAISE Act just require AI companies to publish a safety framework describing how they will test their models for dangerous capabilities and then to follow it. The frameworks are written by the companies, and testing can be done internally. Only Illinois’s SB 315 requires companies to undergo an annual third-party audit starting in 2028.

“There’s a lot of headroom for increasing not only reporting requirements for these companies, but also review by external bodies,” says Peter Salib, a law professor at the University of Houston Law Center. Those reviewers could be private auditors accredited by the government but chosen and paid for by the AI companies. Alternatively, they could be government agencies or insurance companies. 

Legislation

None of this is an accident. The laws on the books that failed to hold AI companies accountable for agentic cyberattacks emerged amid fierce lobbying by the AI industry. 

SB 1047, the California AI bill that was vetoed by Governor Gavin Newsom in 2024 after lobbying by OpenAI, Meta, Anthropic, and the venture capital firm Andreessen Horowitz, proposed a much tougher set of rules. It would have required AI companies to report a broader set of safety incidents (including incidents in which a model acts on its own or slips its controls), undergo annual third-party audits, and maintain a kill switch. But after a year of intense negotiations, Newsom signed SB 53, which narrowed the types of incidents deemed reportable and dropped the requirements for audits and kill switches. 

New York’s RAISE Act followed the same arc. “The version of the RAISE Act that the NY Legislature passed would have required disclosure of this ‘incident,’” Alex Bores, the New York state assembly member who sponsored the bill, wrote on X. New York’s original bill also included third-party audits.

With political pressure mounting, new bills creating better reporting, auditing, and liability regimes for AI development are on the horizon. In Congress, the AI Incident Reporting Act would require AI companies to report to the Commerce Department when a model evades human oversight or breaches a system, even if it doesn’t cause any harm. The Frontier Act would require incident reporting and independent audits. In New York, the Understanding Artificial Intelligence Act, sponsored by Bores, would make companies liable when a model does something that if carried out by a human would be a tort or crime. 

As AI agents increasingly become better at launching cyberattacks, the law remains behind. Closing the gap will require lawmakers to move faster than the next breakout.

The US government wants to spend $30.3 million over the next five years on an improved form of lie detector, according to a Department of Defense budget request. The program, called Polygraph+ or Polygraph Next, will focus on scoring algorithms that use artificial intelligence and machine learning and on a technique called “standoff sensing,” which refers to the ability to take physiological readings without attaching a device to a subject’s person.

According to details of the budget document, which were first reported by Inside Defense, the project will “modernize federal polygraph and credibility assessment technologies” to improve their accuracy and reliability. But the project may be just the latest in a long line of failed attempts to use technology to detect lies. “It’s a misguided effort to reduce the complex to something that is tangible,” says Kyri Kotsoglou, a legal scholar at Northumbria University in the UK who studies the use of polygraphs in the justice system.

The move comes at a time of high tension within the department. Under Defense Secretary Pete Hegseth, the Pentagon has been increasingly turning to polygraph tests in an attempt to find the sources of alleged leaks to the press. In September, the New York Times reported that around 50 officers on the Joint Staff had been given polygraph tests after news coverage reported on the depletion of US weapons stockpiles in the war with Iran. 

Polygraph+ will be run by the Defense Counterintelligence and Security Agency (DCSA), which conducts background checks for the federal government. According to the budget document, which has not yet been approved by Congress, the new technology will be used for vetting of prospective employees and “insider threat detection.” It is not yet clear which specific technologies will be used, and the DCSA did not respond to a request for more information. 

But other Pentagon efforts offer potential clues. In 2023, the department’s Defense Innovation Unit (DIU) ran an open submission process to find companies with products that could be used for deception detection.

It selected two companies to build prototypes: Presage Technologies, which claims to be able to measure heart rate and breathing rate using standard cameras, and Altec Research, a medical sensor company now branching out into non-contact sensing technologies. A screenshot of Altec’s prototype technology released by the DIU shows that it tracks head movement, facial skin temperature, and pore activity. Presage Technologies and Altec Research did not respond to requests for comment. The DIU declined to comment. 

Current lie detection technology has barely changed since the polygraph was invented in the 1920s. Examiners rely on blood pressure, pulse, breathing, and sweat measurements to determine if someone is lying. They make judgments about the veracity of respondents’ replies on the basis of differences in their physiological response to baseline questions like “Is the sky blue?” and target questions like “Have you ever committed a crime?”

The federal government conducts tens of thousands of the tests each year while screening employees, but the reliability of this technology has been repeatedly challenged—and its results are rarely admissible in court. In 1983, Congress’s Office of Technology Assessment concluded that there was very limited evidence supporting the polygraph’s use for screening employees, and in 2003, the US National Research Council (NRC) said evidence for its efficacy was “weak at best.” 

Research suggests humans can spot a lie just over half the time without any technical assistance. The American Polygraph Association claims the polygraph is between 80% and 94% accurate. But the 2003 NRC report pointed out that a screening test with this level of accuracy could still lead to a lot of mistakes. The DOD employs 2.8 million people; an imperfect system applied at that scale could end up falsely accusing tens of thousands.

There are other issues too. Polygraph interpretations are often subjective: Different examiners get wildly different results, and people from minority groups are more likely to be judged as deceptive. What’s more, with training it’s possible for interviewees to learn a variety of countermeasures that can help beat the test; for example, they may artificially heighten their physiological response to baseline questions by stepping on a pin hidden in their shoe. 

“If you know how it works, you can beat it,” says Sophie van der Zee, an associate professor who studies deception at Erasmus University in Rotterdam. She says the machine’s biggest effect is deterrence—often, subjects confess before it even begins. “But that only works if people think a polygraph works,” she points out.

Various new strategies for lie detection have been attempted over the decades, with technologies ranging from thermal cameras to pupil trackers to brain scans. None of them have yielded reliable results outside the lab. The problem is that there’s no single telltale sign of lying that’s true for everyone all the time. “There is still no Pinocchio’s nose,” says van der Zee. 

AI could theoretically improve polygraphs if it could find patterns in the data that examiners can’t. AI algorithms are also more likely to be used for “multi-modal” deception detection, which seeks to combine multiple measurements into an overall deception “score” that is harder for people to game. Three things are happening under the surface that lie detection tries to home in on, van der Zee says: physiological stress, cognitive load, and the conscious efforts people make to conceal the fact that they’re lying. Current polygraph technology tackles only one.

“The more you can have combined methods that approach it from these three different angles, the more successful you will be,” van der Zee says. This isn’t a new concept—in the 2000s, researchers at Manchester Metropolitan University in the UK developed a system called Silent Talker that generated a deception score from video footage. This was later folded into iBorderCtrl, an EU-funded pilot program. In the US, a project called AVATAR incorporated eye tracking, voice analysis, and body movement detection into a tool for use at border crossings. All these projects have quietly faded away. 

Kotsoglou says combining AI and the polygraph is “the worst of both worlds” because it adds uncertainty on top of invalidity. Even if AI or machine learning can spot previously unseen patterns in physiological data, it won’t be able to reliably link them to lying because there is no real ground truth. 

“Even if you have all the records in the world from polygraph tests, you don’t know whether those polygraph tests are right or not,” says Marion Oswald, a professor of law who has written with Kotsoglou on the use of polygraphs in the justice system. She fears that new forms of lie detection will, like the polygraph, be used more as a psychological prop than a scientific tool. 

“It seems very much a response to the concern of the current administration to leaks and perceived lack of loyalty,” says Oswald. “[Lie detection is] being used as a threat, to intimidate and force people to confess to things, as opposed to anything that’s actually getting valid information.”

Around this time last year I was attending an aging conference in Manchester, listening to a talk about fly aging, when my phone started pinging. News outlets were reporting that a hot mic had caught Russia’s and China’s leaders discussing the possibility of living forever.

“With the developments of biotechnology, human organs can be continuously transplanted, and people can live younger and younger, and even achieve immortality,” Russia’s Vladimir Putin reportedly told China’s Xi Jinping.

He seems to have been referring to the “replacement” theory of longevity, which has been supported by multiple experiments that involved physically stitching young mice to old ones. Something about the young blood rejuvenated the old mice. Perhaps young organs could rejuvenate world leaders in their 70s, too.

Unfortunately for Putin, new research pours a little cold water on this idea. Studies on transplanted hearts in both mice and humans suggest that new hearts soon adopt the biological age of the recipient, no matter how young they were to begin with. The finding could be important for transplantation, but it also highlights just how complex aging—and rejuvenation—are.

Jesse Poganik at Brigham and Women’s Hospital in Boston is one of the many scientists trying to understand exactly what it is about the bodies of young mice that rejuvenates old ones. Plenty of research has focused on seeking the secrets of youth in young blood. But what if it’s something about young organs instead?

To find out, he and his colleagues performed a set of heart transplants in mice. In humans, heart transplants typically involve removing a damaged or injured heart and replacing it with another from a donor who is usually much younger than the recipient. (When Poganik assessed hospital records, he found that most recipients were about 20 years older than their donors.)

The mouse transplants were different: Mice received a second heart, implanted in the neck—a procedure that’s slightly simpler and allows scientists to compare the new hearts with the existing ones. In some cases, young adult mice were given a heart from a middle-aged donor. In others, middle-aged mice got young hearts.

The team used a trio of “aging clocks”—molecular tools used to estimate the biological ages of tissues and whole organisms—to assess whether the additional hearts affected the mice in any way. These clocks were good at predicting the chronological age of mice that didn’t get new hearts.

Poganik says he was expecting to see a reciprocal effect, and that young hearts might benefit older animals, for example. In previous work by other members of his team, young mice that got old hearts experienced a buildup of senescent cells in their other organs. These cells are thought to contribute to the aging process, suggesting that receiving an old organ might prematurely age an animal.

But that’s not what he found. When he and his colleagues analyzed the transplanted hearts between four and six months after surgery, they found that the hearts seemed to have adopted the biological age of the recipients. Young hearts got older, and old hearts got younger. “The environment of the transplanted organ really dictates how it seems to behave biologically,” he says. The findings were published online at bioRxiv last week.

The team also looked at each mouse’s blood and other organs—including its original heart—and were surprised to find that they seemed to be unaffected by the presence of the new heart, despite the age of its donor.

The finding was backed up by data from human heart transplants. People who receive a donor heart must typically undergo a series of heart biopsies after surgery. The tiny pieces of heart tissue collected from people who had heart transplants at Brigham and Women’s have been stored for decades. And when Poganik and his colleagues tested their biological ages with the aging clocks, they found a similar pattern: No matter the age of the donor, a new heart quickly adopts the biological age of the person who received it.

It’s not clear why this is, but João Pedro de Magalhães, who studies aging at the University of Birmingham in the UK and was not involved in the study, thinks it might have something to do with the recipient’s immune system. Perhaps immune cells circulating in the blood might affect markers of aging in the new organ, he says.

Perhaps the potential rejuvenating effects of a young heart end up being diluted by all the other aged components of an older body, Poganik suggests. Or maybe a single organ just isn’t enough to see an effect.

Poganik hopes his finding will encourage surgeons to consider using hearts from older donors, many of which are discarded on the assumption they won’t function as well. (Machines used to preserve organs before transplantation are changing that already—and Poganik has worked on another study showing that these devices seem to rejuvenate donor livers to some extent.)

But the study also highlights just how complicated aging is. If organs seem to be getting older by one measure but not by another, how can we get a full picture of the biological age of an organ, or a person?

This complexity means that scientists are not likely to discover a true way to completely reverse aging, says Poganik. “That would mean that every aspect of aging has to go back in time,” he says. Reversing DNA damage, structural damage, and all the other degradations that are part of the aging package, across all our various cells and tissues, presents an enormous challenge.

“There are aspects of biological age that are probably reversible, and there are aspects that are probably not,” he says. Sorry, Putin.

This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

This week, world leaders descended on Manhattan for the UN General Assembly. It’s also New York Climate Week—investors, policymakers, advocates, and journalists are colliding at panels, talks, and fancy dinners.

With so many climate voices in one place, the discourse can feel a little louder than usual. This year, the unavoidable topic is artificial intelligence.

There’s been a growing tension bubbling up about AI’s impacts on climate and climate tech. Depending on where you stand, you might point to AI’s potential for advancing research, or to the way funding and attention from Big Tech is trickling into energy startups. Or you might focus on the emissions-heavy natural-gas buildout that’s unfolding to meet the sector’s electricity demand.

Everyone seems to agree that AI is important. The big question for those in the climate world, and the conversation I’m constantly having and hearing this week, involves how you see its influence unfolding.

UN Secretary-General António Guterres highlighted the AI and climate crossover in a speech on the first day of the assembly. “The climate crisis fuels instability and displacement,” Guterres said. “Artificial intelligence could help solve all these challenges, or it could make them worse.”

It feels relevant that this year is the first that the world has really had to grapple with the fact that climate goals are slipping out of reach. A recent report from the UN Environment Program said that the world has nearly passed the point where we could possibly keep warming to less than 1.5 °C above preindustrial levels.

Since we’ve essentially missed this target, the report lays out the need not only to quickly and drastically reduce greenhouse-gas emissions, but also to employ carbon removal to help suck up emissions that have already been released into the atmosphere.

The billion-dollar question is whether AI could help with any of this. 

The energy-intensive technology is certainly shining a spotlight on the need to build out electricity supplies and shore up grid reliability. The result is more attention and money for energy technologies, some of which happen to be low- or zero-emissions.

As I’ve covered before, startups across the energy and climate sectors are benefiting. Firms in nuclear, geothermal, wind, and solar power have signed deals with the likes of Google, Meta, and others looking to power their new or growing data centers.

Global climate-tech investment from venture capital hit $26 billion in the first half of 2026, according to data from Currence, a finance tracker for the industry. That’s 55% higher than last year, and products and services for data centers are getting a massive slice of that pie.

But as a Semafor piece about the report points out, some sectors are slipping through the cracks: Carbon management and low-carbon fuels saw VC investment plummet this year. These are important solutions for addressing climate change but may not be able to sell themselves to a data center.

So far, the data center buildout has come with a hefty emissions toll. A few years ago, Microsoft, Google, and Meta all had ambitious goals to reduce greenhouse-gas emissions. Now they’ve all seen emissions rise, largely because of data centers that are needed to power AI.

Some people are optimistic. AI could help speed up progress in areas like the search for new catalysts, as Evelyn Wang, MIT’s VP of energy and climate, pointed out during a panel. And data centers won’t add to climate and water problems forever, Wang told the Associated Press. (She puts the timeline at about a decade until data centers no longer add to planet-warming emissions.)

But a whole lot of natural gas is coming online to meet the immediate demand created by new data centers. And once those power plants are built, they have a decades-long lifetime. 

That’s partly why public pushback to AI is growing—people are seeing more pollution and noise near these data centers and the power plants that provide them with electricity. Overall, what I’m hearing this week is that many in the climate sector are skeptical of AI, at best. 

“AI leaders are now on thin ice when it comes to license to operate and sinking deep underwater when it comes to public support,” said UN climate chief Simon Stiell in a speech this week. “Tech titans need to start showing why the benefits of AI outweigh its skyrocketing costs—for the many, not just the tiny few.” 

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

Brace yourself: It turns out AI is being optimized for cheating. OpenAI’s agents hacked into Hugging Face to get the answers to a cybersecurity test. Next, they solved a prestigious math problem (or just stole from two top mathematicians’ answer sheets). Anthropic’s models have also hacked into other companies’ systems four times already. And that’s only what we’ve caught so far. 

Freaking out? You’re not alone. AI lab researchers are quitting their jobs and issuing dire warnings that if we keep going this way, AI might eventually kill us all. Bill Gates is sounding the alarm. Bernie Sanders has teamed up with Steve Bannon, of all people, to call for curbs on AI. Anthropic CEO Dario Amodei is urging a slowdown, and other top US AI executives agree. But fear not: President Trump has a plan. He says the only guardrail AI needs is “a STRONG AND SMART (High IQ!) PRESIDENT.”

Delia Ramirez, a Democratic US representative from Illinois, has announced a plan to introduce new legislation to terminate the surveillance tower program along the US southern border. 

The announcement comes just days after publication of an MIT Technology Review investigation, “Dying on Camera,” in which we looked at deaths along the border that took place close to one or more of these towers. MIT Technology Review found that nearly one in four deaths we analyzed between 2015 and early 2026 occurred within their advertised range. We also showed that the Department of Homeland Security does not keep track of these deaths or other metrics on the towers’ success or failure. 

Ramirez, who sits on the Homeland Security Committee, cited MIT Technology Review’s reporting that nearly “1,100 people have died within range of a billion-dollar surveillance tower system between 2015 and 2026,” adding, “The towers that we have paid a billion dollars to just don’t work.” 

There has been growing public support for abolishing ICE in the wake of the second Trump administration’s highly visible enforcement operations across American cities, the buildout of its detention apparatus, and shootings by ICE and CBP agents. Wilber Rafael Garcés Pérez, a delivery driver, was shot in the back by an ICE agent just last weekend in Austin, Texas. 

But for Ramirez, abolishing just ICE is not enough. At a press conference held in Chicago on Wednesday, Ramirez said, “There’s no question that DHS has to be dismantled. We need to provide them a road map in how we do it,” Ramirez said. 

It will not be the first time that a bill to dismantle the Department of Homeland Security has been introduced, but previous proposals have not gone very far. This proposal will also face a steep uphill battle in Congress, where even some Democrats favor reforming rather than outright dismantling the organization.  Ramirez, however, sits on both the House committee that oversees the department, and is the ranking member on the Homeland Security Committee’s cybersecurity subcommittee.  

“We’re limiting state surveillance, we’re reigning unlawful abuses of data and curbing the militarized enforcement. We’re beginning the process of establishing a separate civil immigration system,” Ramirez said.

CBP did not offer an immediate response for comment. Markwayne Mullin, the Secretary of Homeland Security, said, “Radical leftists are calling to defund the police and abolish DHS. It’s NEVER going to happen.”

This story has been updated to include a response from the Secretary of Homeland Security.

Shubnam was packing boxes for a move into a new home when their friend sent them an Instagram video. The footage had only been up for a few hours, but it was days old, recorded at a Delhi protest this spring against a bill that would have narrowed the legal recognition for transgender people in India. 

There, a content creator known for rage-bait videos had approached Shubnam, a 38-year-old transfeminine graphic designer and educator. Shubnam, who uses a single chosen name, was dressed in a green sari with a shaved head, and the creator asked why they were wearing “female clothing.” Shubnam told him to back off. Only when the video arrived did they realize he had recorded the exchange on Meta smart glasses, styled to look like ordinary Wayfarers. 

Edited into a mocking reel, the footage drew millions of views, along with transphobic abuse, reaction videos, and AI-generated memes. It spread from Instagram to X and YouTube. Copies remain online today. 

For Shubnam, the video has caused huge harm. “All it took was one unconsented video to rewind two decades of progress that I made all on my own,” they told MIT Technology Review. “For a moment, it felt like this entire adult life I had lived was just a long lucid dream, and I was going to wake up as that clueless queer child again.” They haven’t worn a sari outside their neighborhood since.

Experts warn that many others will experience similar ordeals as smart glasses go mainstream. And they say privacy problems are becoming especially acute in India, where covert recording and the circulation of images without consent are already pervasive.

Cases of cyberstalking, sextortion, deepfakes, and other forms of online abuse are growing in the country. And the chances of a crackdown on technologies such as smart glasses are slim, given that Indian law enforcement authorities are themselves increasingly adopting AI-powered surveillance tools, including facial recognition systems and their own AI-enabled smart glasses.

An unequal burden

Smart glasses are among the fastest-growing devices in India’s wearables market, according to the International Data Corporation, a research firm. And Reliance Jio, one of the country’s largest telecommunications companies, says that later this year it’s set to launch a sub-$105 rival to Meta’s glasses, the latest model of which have a $420 price tag in India. 

The core privacy issue is that these cameras don’t look like cameras, so targets may not notice they’re being recorded. Yet Meta rejects the idea that it’s responsible for protecting people from intrusions, says Janusz Świerczyński, an Oxford Saïd Business School researcher who studies privacy and emerging technology.

“More and more, the expectation is being placed on bystanders to look out for themselves,” he says. That means it’s up to them to spot a small LED in the glasses frames that activates when recording is in progress. Meta also relies on wearers to behave responsibly, says Świerczyński. Reliance Jio did not respond to our request for comment on privacy issues related to its upcoming glasses. 

Creators have already demonstrated ways to circumvent the recording light on Meta’s glasses. Tutorials show users how to cover or otherwise obscure the LED while continuing to record. Back in March, Wired documented a market for “stealth mode” modifications that physically disable the light or remove it entirely. 

All that makes covert recording easier, and the risks are not borne equally. Świerczyński says women and other marginalized groups face greater harm when inconspicuous devices capture and circulate footage without consent—as happened to Shubnam.

The bigger problem, however, is that in India smart glasses aren’t just being used to turn ordinary people into targets of viral “pranks.” They’re also being adopted as a tool of police surveillance. 

In June, when thousands of young demonstrators gathered to protest problems with the country’s education system, Delhi police used Meta smart glasses to record them. In fact, a petition filed by former Jawaharlal Nehru University Students’ Union president Aishe Ghosh before the Delhi High Court alleged that police had filmed protesters continuously for weeks, including while they were eating and resting. It also alleged that the police had threatened to send footage of student demonstrators to their parents and colleges.

In July, India’s solicitor general dismissed the case in court as “luxury litigation.” Instead of investigating the students’ complaints, police turned on the students themselves, opening 10 criminal investigations for offenses including rioting, assault on public servants, and damage to public property. Some of the protesters, who spoke to MIT Technology Review on condition of anonymity, say they never saw the LED on the Meta AI glasses police wore during the demonstrations. 

Filming without consent

Meta maintains that its safeguard works. The company told MIT Technology Review, “Every pair has a capture LED that blinks when you take a photo or video that you can save or share; it can’t be turned off, and if someone covers or damages the LED, the camera is disabled.” The company also says users are responsible for complying with the law and respecting people’s privacy.

That approach is risky, especially in India, where norms around filming and consent are still highly disputed and digital literacy is still uneven, says Sameer Parmar, a counselor at Meri Trustline, a safety partner working with Meta and YouTube to support people harmed by online content. 

“The hardware is improving, but the country’s weak understanding of consent gives an invisible camera far more room to cause harm,” he says. Simply being in a public place is often treated as implicit consent to be photographed or filmed. He says the help line has been seeing an uptick in cases involving covert recording and nonconsensual footage.

“Violations of this access have increased significantly and are expected to rise further in India,” he says. People are recorded in medical settings, during private encounters, and in public spaces. Those with less power to object are particularly vulnerable, he adds.

“It’s either trolling or it’s playing pranks on them and then recording them” he says. Abusers “do target people who are a little more unaware or susceptible to how this technology-facilitated harm works.”

The law in India has not caught up with this new kind of harm, says Apar Gupta, a lawyer and founder-director of the Internet Freedom Foundation. “India’s laws against voyeurism and the sharing of intimate images were designed largely around more recognizable forms of abuse,” he says, citing offenses such as secretly filming someone undressing, using a toilet, or engaged in a sexual act. “[The laws] offer limited protection when someone is secretly filmed in an ordinary public setting and the footage is then used to humiliate them,” he says.

Social media platforms’ rules do offer another route to removing harmful content filmed with smart glasses, but it’s not an immediate remedy. “Under India’s current rules, platforms can have up to 36 hours to act on certain complaints—a window that can be long enough for a video to be downloaded, copied, and circulated elsewhere,” he says.

The problem of filming without consent seems set to grow. According to the Financial Times, some internal prototypes of new Meta glasses include AI features that continuously photograph and listen to the wearer’s surroundings, enabling the glasses to later recall what they saw or heard. These functions wouldn’t activate the external recording indicator, making it impossible for bystanders to know when information about them is being collected.

Meta told us it intends to keep the LED for active capture, such as taking photos or recording video, but not for AI features that interpret the wearer’s surroundings. Its argument is that a constantly blinking light could eventually become so commonplace that people stop noticing it.

But that logic risks putting society in a perpetual game of catch-up, says Woodrow Hartzog, a professor of law at Boston University. “Tech companies have an incredible ability to push these tools out faster than society can meaningfully acclimate to the threat,” he says. “By the time society catches up, there’s already a degree of normalization.”

For Shubnam, the danger is not just that people may become accustomed to being recorded. It is that they will become accustomed to seeing versions of themselves that they never chose to put out into the world. 

“Being comprehended along the lines of a narrative that I didn’t compose or consent to—that imposed recognition makes you feel like you’re powerless in how people know you,” they say. 

It’s been a busy few months for AI hype. At the end of April, Anthropic claimed that its model Claude Mythos is better at finding software vulnerabilities than most security experts. Then we had the OpenAI–Hugging Face hacking incident, after which Anthropic (proudly) and Meta (reluctantly) disclosed similar incidents involving their models.

This was followed by Anthropic’s claim that one of its models had made a mathematical breakthrough; soon OpenAI claimed a mathematical breakthrough of its own. Most recently, Anthropic engineer Jacob Coxon went viral announcing his departure from the company, claiming that it and OpenAI are “racing straight towards self-improving superintelligence and gambling with our lives.”

Each of these events was mostly covered breathlessly by the press, often repeating the companies’ anthropomorphizing framings—which are designed to portray their software is not only powerful but incipient “artificial general intelligence.” So what is really going on? Are we witnessing a massive, civilization-changing set of technological breakthroughs, or is this marketing?

In all these incidents, massive fanfare from the companies (presented as mea culpas in illicit hacking cases) is accompanied by intense press coverage. Once there is time for experts in the relevant fields to examine what happened, a very different story emerges, but one that gets less media attention.

Regarding the “hacking” incidents, cybersecurity experts say the story is more about OpenAI’s negligence and failure to adopt basic, established security practices than about “models gone rogue” or “AI agents creating civilizations.” 

As for the mathematical results, mathematicians who were initially “stunned” by OpenAI’s press release saying that its latest chatbot, Astra, solved problems that “have been open and seen no progress on the main result for at least a decade”—but they later realized that the results weren’t as “novel as first appeared.” Since then, mathematicians have accused the company of research misconduct and plagiarism, and they’ve reiterated that Astra didn’t make a “profound intellectual leap.” Just weeks later, OpenAI claimed its own mathematical breakthrough. Two days before, Tristan Buckmaster, a math professor at New York University’s Courant Institute, published a bombshell statement suggesting that OpenAI had stolen other people’s work and improperly attributed it. 

Claims of incipient, dangerous superintelligence are not based in good scientific or engineering practice. Rather, they are narratives based in ideologies of transhumanism, eugenics, and wishful thinking about imagined future digital humans.

It’s worth thinking about why there is so much attention on computer programming and math as fields in which to apply large language models and related technology. Not only are they often elevated as the pinnacle of human intellectual achievement, but they involve problems where answers, once suggested, can be verified. The former property helps AI hype mongers sell the idea that they are building everything machines. The latter makes math and coding problems easier to tune systems for, since system output (sequences of likely words or pieces of computer code) can be evaluated without having to pay data workers to look at and annotate each one.

Mathematicians in particular have warned against corporations using their field in this way. A statement signed by hundreds of them says there is “currently a strong commercial incentive on the part of the technology industry to overstate the capabilities of their products” and asks policymakers to “consult with experts, including mathematicians, in forming policy decisions rather than relying on press releases or popular reporting of mathematical results.”

We echo this call and note that the illusion of speed and urgency promulgated by the tech companies is also a ploy to misdirect both policymakers and the public. Unfortunately, it sometimes works, such as with Senator Bernie Sanders’s well-meaning but ultimately misguided proposed legislation to prevent the development of “artificial superintelligence.”

Describing them as “superintelligence” or “rogue models” ascribes agency to products rather than to the companies building them. This framing markets these companies’ products as “superhuman” and, at the same time, helps the companies evade accountability for their actions. 

Instead of OpenAI being prosecuted for creating malware that hacked another company, press releases, news outlets, media personalities, and lawmakers refer to “rogue models” as if they acted on their own. Instead of researchers being questioned about their companies’ habit of plagiarizing academics’ work or using customer data to train models without consent, the public’s imagination is redirected to fears about what the future might hold upon the arrival of fictional superintelligent machines.  

The AI industry has even suggested that popular, bipartisan anti-data-center activism is a “distraction” from attempts to regulate the impending, scary, “superhuman” machines these companies are building. According to the AI industry, we should be more worried about a fictional machine god than about the climate catastrophe that these data centers exacerbate, the asthma suffered by those living near them, the rising electricity bills of the public subsidizing them, or the water that is redirected to cooling them. 

We know better than to make decisions based on marketing and better than to capitulate to corporate pressure to make those decisions quickly. Wise decision-making, by policymakers and communities, demands time to hear from independent experts and contextualize corporate claims. The best possible outcome from this summer of hype is that policymakers and the public at large learn to take a breath, hold onto our skepticism, and recognize this kind of hype for what it is the next time it comes around.

Timnit Gebru is executive director of DAIR and author of the forthcoming book Deep Unlearning: The Radicalization of a Tech Idealist, which is available for preorders now and set to publish on February 16. Emily M. Bender is professor of linguistics at the University of Washington and coauthor of The AI Con.

Our 15-month investigation into death and surveillance along the US-Mexico border began with a simple question: Why did so many people die near government surveillance towers meant to help track and apprehend them?

The result of our investigation is the first comprehensive map and analysis of deaths near border surveillance towers. Here’s how we built it, what decisions we made along the way, and what the data can and can’t tell us.

Our data

The investigation relied on knowing where migrants have died and where and when US Customs and Border Protection towers were installed. We analyzed cases dating back to 2015, allowing us to cover different border policies, presidential administrations, and tower technologies.

Migrant deaths

The US-Mexico border has been called the world’s deadliest land border, and Border Patrol estimates that more than 10,000 people have died during crossings since 2000—a figure that’s widely considered an undercount. But the amount of information available publicly on where and when these people died, and who they were, varies widely. Some remains are never discovered, and for those that are, there’s no national protocol for how the records are handled.

We considered a case for further analysis if we could confirm three basic things: whether the person was believed to have been crossing the border, where the remains were found, and roughly when the person died. 

We created our dataset by merging existing ones that had been compiled and shared by other organizations, including No More Deaths, Humane Borders, and the Electronic Frontier Foundation, with new records we obtained ourselves from more than a dozen agencies.

Public records requests in Texas

Texas was the biggest missing link in most existing databases of migrant deaths. Unlike Arizona, it has no initiative to share records online, and unlike New Mexico, it leaves individual counties to handle their own death investigations. Some records for those individual counties existed, compiled by a handful of dedicated researchers, but nothing was comprehensive or current. 

We identified 17 counties in Texas that would be relevant to our investigation: Culberson, Jeff Davis, Presidio, Terrell, Val Verde, Kinney, Maverick, Dimmit, Webb, Zapata, Jim Hogg, Starr, Hidalgo, Cameron, Kenedy, Brooks, and Duval. There are other counties in which migrants have died, but they do not have areas covered by surveillance towers and were therefore excluded.

In most of these counties, locally elected administrative judges called justices of the peace collect the most comprehensive information on local deaths. But records requests submitted to these individual judges can stall for years (indeed, some of our requests to justices of the peace remain unfilled more than a year after we filed them). Reports from justices of the peace sometimes leave out the scene or location details we needed; some justices never even visit the place where a migrant died, instead formally declaring the person deceased over FaceTime with the first responders on scene.

MIT Technology Review instead filed records requests to the sheriffs of these 17 counties beginning in July 2025. These agencies are often first on the scene when a death is discovered, and their reports can have more detailed information on where remains were found. But relying on sheriffs’ records means our analysis undercounts migrant deaths in Texas, as it doesn’t include any deaths handled by local or state police.

After we filed a request, it took periods of near daily calls to the offices to receive the records, if we received them at all. Some were provided with no charge. Other counties charged between $300 and $1,300 to fill our request. Counties warned that their records were incomplete, with an unknown number lost during a move, damaged, destroyed, or misplaced during transitions between sheriffs.

Those delays were compounded because most agencies do not record whether a person is believed to have been crossing the border when they died. That’s despite provisions in some counties that allow officers to track far more granular details about other situations; Zapata County reports have a checkbox to indicate whether jewels were stolen in a burglary, for example, while Cameron County reports have one for whether a burglar entered through a chimney. Without a way to readily identify migrant deaths, offices had to pull records by hand, making the process slower, more costly, and more prone to error.

We received usable records from 14 of these counties (records have yet to be received from Dimmit or Duval, and Maverick’s office charged a per-record fee that was cost prohibitive).

Agencies generally sent us police reports for each individual case, amounting to over 4,000 pages of records in total. Some were handwritten documents. 

Except for the reports from Kenedy, Webb, and Hidalgo Counties, we pulled out the relevant information by hand. For those three counties, which sent large volumes of records, we used Anthropic’s Claude, accessed via API, to inspect each case report and pull out coordinates of the spots where remains were found; then we checked batches of those cases by hand to verify the AI’s accuracy. Agencies sent us records on a rolling basis from July 2025 to February 2026, and most did not specify the date through which their records were current.

We generally trusted that agencies sent us cases they believed involved migrant deaths, since they had the most information about each case. But some records clearly suggested otherwise (indicating, for example, that someone died at home), and we excluded those cases.

Several hundred cases were removed from our analysis because they didn’t include enough information for us to be certain where the remains were found. In fewer than 100 cases, coordinates were not listed but the description was specific enough for us to locate a point within 0.25 miles of where they were found—so we kept those in. 

Compared with agencies in other states along the border, those in Texas provide far fewer details regarding how long ago someone may have died before remains were found or how far the decomposition of those remains had progressed. Some reports included this information, but many didn’t. While autopsy records sometimes have more details, many migrant autopsies in south Texas are done through the Webb County Medical Examiner, which charges a fee for each report. That was cost prohibitive given the scale of our analysis.

MIT Technology Review consulted with Greg Hess, director of Arizona’s Pima County Office of the Medical Examiner, on a framework for translating descriptions of the scenes, bodies, and causes of death available in many of these records into rough estimates for when a person may have died (more detail on this process is provided below).

In total, we included more than 1,500 cases from our Texas records requests in our final analysis.

Accessing Arizona records directly

The Pima County Office of the Medical Examiner (PCOME) in Arizona handles death investigations for all border counties in the state except Yuma (we sourced Yuma’s records from the organization No More Deaths, as detailed below).

The office determines which cases it believes to be border crossers. It publishes a data portal for these migrant deaths that’s updated monthly and shares its data with the organization Humane Borders, which then publishes it in a map and database. We included the data from 2015 through April 2026.

Humane Borders notes for each case how precise the location description is. We included only cases with GPS coordinates precise to within 300 feet. Collecting these coordinates has been standard practice since around 2012. The records also include an estimate of when the individual died, ranging from less than a day to at least six to eight months before remains were found.

In total, we included more than 1,700 cases from Humane Borders in our analysis.

Data collected by No More Deaths

To obtain data on deaths in California, New Mexico, Arizona’s Yuma County, and El Paso and Hudspeth Counties in Texas, we turned to the nonprofit organization No More Deaths, which has tracked migrant deaths since 2004. It publishes information about its records requests and its methodology. 

Many of the cases in its database include GPS coordinates, sometimes with notes that the location is only approximate—for example, accurate to within half a mile or one mile. We included only cases where the location was known to within a quarter-mile.

No More Death’s database includes, when available, information on the level of decomposition in which each set of remains was found. As with cases in Texas, MIT Technology Review consulted with PCOME’s Greg Hess on a framework for translating this information into rough estimates for when a person may have died. 

Here’s where the No More Deaths data used in our analysis comes from:

• In California, No More Deaths requests records from the San Diego County medical examiner and the Imperial County coroner. Its latest data for Imperial County is from December 2025, and its latest for San Diego County is from October 2025. 
• In Yuma County, Arizona, No More Deaths requests records from the Yuma County medical examiner. The latest data is from September 2025.
• In New Mexico, the organization requests data from the state’s Office of the Medical Investigator. Unlike most states, New Mexico has a statewide medical examiner, allowing No More Deaths to obtain migrant death data for the entire state. The latest data is from July 2025.
• For Hudspeth County, Texas, No More Deaths requests records from justices of the peace in Districts 1 and 2. The latest data is from December 2023.
• For El Paso County, Texas, records were from the El Paso County Office of the Medical Examiner. The latest data is from September 2025.

In total, we included nearly 1,000 cases from No More Deaths in our analysis. 

Total cases reviewed

When accounting for all our data sources and excluding deaths without specific enough records to determine where remains were found, we analyzed a total of nearly 4,000 cases. 

Other notes 

We generally assume that a person died where their remains were found. That may not necessarily be true: Remains can be moved by other people, flowing water, or animals. But medical examiners and researchers we consulted said such cases are rare.

We excluded cases in which the virtual wall did not appear relevant. For example, we removed cases involving people who died inside Border Patrol stations or in car crashes during pursuits. 

The deaths we analyzed reflect a variety of causes that may seem at first like very different sorts of surveillance failure. If someone slowly died of dehydration within view of a camera, for example, agents would have had far more time to intervene than they would have in a case where someone drowned quickly in a river or canal. But we included all these deaths because, in each one, the virtual wall could have prompted a response from agents.

Surveillance towers

Including a tower in our analysis required knowing three things: its location, its type (so that we could judge how far it is supposed to see), and the time it was installed. 

Locations

The Electronic Frontier Foundation has been tracking and mapping CBP surveillance towers since 2023, using on-the-ground reporting, satellite imagery, and government documents obtained through public records requests. Its tally is an undercount; government records suggest that about 800 towers are currently deployed, while EFF has logged about 600. Our analysis therefore misses deaths that occurred near towers not yet catalogued by EFF.

MIT Technology Review worked closely with EFF’s director of investigations, Dave Maass, throughout this project. Our tower data began with the database published by EFF in April 2026, but that map did not include towers that have been removed. We consulted EFF for the details on these former towers and added them to our database. We generally use EFF’s reference numbers for individual towers, but we changed some to ensure that each tower has a unique identifier.

The locations are identified by GPS coordinates. We verified each tower’s coordinates by analyzing satellite imagery.

Tower types

Our analysis focused on towers of three main types. Certain towers mapped by EFF don’t fall into any of those categories, so we excluded them because it’s difficult to obtain information about how far they are supposed to be able to see. This includes some towers that are installed at Border Patrol checkpoints or stations.

CREATIVE COMMONS

CENGIZ YAR FOR MITTR

To know how far the main tower types could see, we consulted materials from tech companies and EFF’s research, led our own review of government documents, and interviewed former Border Patrol agents and officials. The distances described are consistent across these sources. But they are advertised estimates, not contractual guarantees: Some government records redact more specific information about a tower’s surveillance range in particular locations, citing security concerns.

When towers were present

To know if a tower in place now was present when someone died, it’s essential to know not just where it’s located but also when it was put there. To figure this out, we used multiple sources of satellite imagery. 

Maass, at EFF, had previously analyzed some of the towers to determine when they first appeared. MIT Technology Review checked satellite imagery for these towers to confirm these timelines, and then checked for all the remaining towers. For each tower, we logged the date that it first appeared and, if it eventually disappeared, the last date it was present. We spent months meticulously checking when hundreds of towers appeared in multiple sources of satellite imagery to confirm there were no contradictions. 

This is easier for certain tower types than others; towers from Anduril, for example, tend to go up where there was no prior infrastructure, and it’s fairly easy to see how a patch of empty desert gave way to the new tower and its trademark solar panels. Other towers, like IFTs, were often built on or near existing structures. For these, we leaned more closely on EFF’s expertise in analyzing the imagery. 

When imagery wasn’t sufficient, we used Google Street View, government documents, or photographs collected by EFF to verify when towers went up.

One limitation is that historical imagery of towers along the border is not available for every day in the past. Especially for older towers or those in remote locations, images may have been taken months or even a year apart. A tower visible in May and again in December could theoretically have been removed and reinstalled in the intervening months. But most tower systems are permanent structures. And multiple Border Patrol agents told us that even towers that are meant to be mobile, like autonomous towers, are in practice very rarely taken down or moved (consistent with this, we observed that semipermanent fences were erected around the sites of many Anduril towers).

Another limitation is that even if a tower is visible at a particular time, that does not mean it was online. We can’t verify that it was functional, receiving power, or transmitting data. Our analysis therefore establishes when a tower was present, not whether it was operational.

Heights

We conducted a topographical analysis, detailed below, to see whether a tower’s view of a given person might have been blocked by the terrain. This required estimating how tall each tower is. 

We included a minimum and maximum height for each tower, based on research from EFF. In select cases where the surveillance system was not on a typical tower but mounted to an existing structure, like a water tower or building, we used satellite imagery to estimate its height. The higher a tower is, the farther it can see.

Our analysis

With the data we collected, we could begin to see which deaths happened in range of the virtual wall. But for a given match between a set of remains and a tower to count, it needed to satisfy two criteria:

1) Distance: The death must have happened within the published surveillance range of the nearby camera.

2) Timing: The tower had to be present when the person was estimated to have died.

Distance

Minho Kim, an outside collaborator who is now a postdoctoral researcher at Stanford with a focus on wildfire mapping, led the distance analysis. His software ingested our database of human remains and our database of towers. Then it created a row for each time a set of remains was found near a given tower, along with the distance between the two.

We then checked whether, for each match, the tower was actually present when the person died.

Timing

Sometimes, as with the Arizona cases, an estimate of when someone died is given in the original data; the medical examiner may note that a person died more than three weeks but less than five weeks before remains were found, for example. 

In other cases, we used the level of decomposition, the cause of death, or both to calculate a window of time in which the person would likely have died. Finally, we compared this window with the time when we determined the tower was built. 

This step took us many months and involved reviewing original police reports for hundreds of cases, inspecting photos of bodies, cross-referencing cases with news articles, interviewing medical examiners, and researching decomposition timelines. The resulting death windows are approximate, but they are the best estimates we could make from the information available in each case.

We assigned different confidence levels to cases based on how these dates lined up. If a person’s entire estimated death window fell well after a tower was first observed, we felt confident the tower was present when they died. If any part of the death window fell before the tower was first observed, we excluded the case.

Many cases involve skeletal remains. Remains decompose differently depending on location and season, but when records didn’t provide a specific estimate, we conservatively estimated that skeletal remains had been there for at least six months. For these and other cases without a specific estimate of when the person died, we used three years as a conservative cutoff: If a tower had been present for at least three years before the remains were found, we included the case, judging it unlikely that the person had died before the tower was installed and remained undiscovered that long. If the tower had been present for less than three years, we excluded it.

Viewshed

It is possible that someone who died within range of a tower could have been blocked from the tower’s view by terrain, vegetation, or buildings.

We estimated the impact of this issue with a technique called topographical analysis.

Kim, our collaborator, led this part of the analysis. Here’s how it works: Imagine a map of the border region divided into a grid, with each square having one value for that area’s elevation. Take a given match between a set of remains and a tower and draw a line between the tower’s estimated height and an estimated height of 5’ 6″ for the person who died. If the terrain along that path rises high enough to block the line between the two points, we estimate the tower’s view as obstructed. If it doesn’t, we estimate the view as clear.

It’s not perfect. In the elevation data we used, each square represents an area 30 meters wide, or about 98 feet, and is assigned a single elevation. That means smaller changes in the terrain can be lost, particularly in steep or rugged areas where elevation changes quickly. Our analysis is therefore less about re-creating exactly what a tower could see than providing a rough estimate of whether terrain may have blocked a tower’s view of the spot where someone died. 

We include results on our map whether a clear line of sight was estimated or not. Even if the location where someone died was blocked from a tower’s view, the person may have passed through areas visible to the tower before reaching that spot.

It’s much harder to estimate whether a line of sight was blocked by vegetation. We are less concerned about this limitation because virtual-wall towers use thermal imaging, which can detect body heat through light vegetation. Dense vegetation, however, can still obstruct their view.

Buildings pose a greater challenge. We could not reliably account for their locations and heights. Instead, we reviewed each tower’s surroundings and flagged towers located in or near urban areas. These notes warn readers that buildings may have limited what those towers could see.

Beyond the data 

The most fundamental question in each of these cases, and for the families of anyone who died, is whether the US government saw someone attempting to cross into the country and never intercepted them—or never saw them at all. 

This is all but impossible to answer with records or data—Border Patrol doesn’t disclose which alerts came in surrounding any given migrant death, and footage isn’t retained unless an internal investigation is triggered, which doesn’t happen unless a migrant dies while in custody.

MIT Technology Review instead conducted dozens of interviews with people who worked for Border Patrol, Customs and Border Protection, or the Department of Homeland Security or who served as presidential advisors. These conversations yielded previously unreported details about how agents use surveillance technology, turned up instances where it has underperformed, and underscored how seldom its effectiveness is evaluated internally.

We also sought out friends and family members of people who died near surveillance towers, using additional records, social media, and other research tools to find them. We approached these interviews using best practices for trauma-informed reporting.

Beginning in September 2025, MIT Technology Review requested interviews with current Border Patrol agents and leadership at CBP, and we asked to visit the border to understand the strengths and weaknesses of the latest autonomous tower program. Following delays, which were in part related to DHS funding lapses, in July 2026 we were told our media request had been approved at the local level in Big Bend, Texas. We wanted to visit this area because Border Patrol is expanding the number of Anduril towers there. But when the request went up to the DHS headquarters for a final sign-off, the agency denied it. Before publishing our investigation, we sent detailed questions to CBP, Anduril, General Dynamics and Elbit. CBP and Anduril responded, but did not substantively address our questions. General Dynamics referred us to CBP. Elbit did not respond to our questions.

MIT Technology Review thanks the following for their assistance with this project:

Minho Kim, Stanford University

Greg Hess, Pima County Office of the Medical Examiner 

Stephanie Leutert, University of Texas at Austin

Don White, Brooks County Sheriff’s Office

MIT Technology Review today published our investigation into how many people have died near the “virtual wall” of surveillance towers that the US government has installed along the US-Mexico border.

We found cases of people who walked undetected through areas surveilled by advanced, AI-enabled towers and later died nearby, where their bodies remained unnoticed for weeks or months. These cases reflect systemic flaws in the virtual wall: broken towers, algorithms that failed to detect people, and agents who simply did not respond to alerts.

Each of these deaths should have raised alarms inside US Customs and Border Protection (CBP) about the flaws in the agency’s surveillance network. This is particularly true for cases near the latest towers built by the defense company Anduril, which continues to win major contracts and is poised to benefit from the windfall of federal money slated for border security. The US is set to spend $1 billion to triple the size of the virtual wall by 2034.

Here are four ideas for what CBP should do next to address these failures. 

1. Conduct a comprehensive audit of deaths near the virtual wall

“If I were the federal government, I would do an audit, because this feels like it’s not working.” 

That was New Mexico state representative Sarah Silva’s reaction when MIT Technology Review showed her that in 2024 alone, more than 18 people had died near surveillance towers in just one small stretch of desert near her district.

An audit is necessary because, despite poring over thousands of pages of police reports and other documents about surveillance towers, we have only limited knowledge of how many towers are out there and where they’re located. 

CBP does not publish a comprehensive map of its virtual wall. We instead relied on mapping work led by the nonprofit Electronic Frontier Foundation and then used satellite imagery to verify where specific towers were and approximately when they were installed. 

We analyzed lots of towers—nearly 600. But that’s still far less than the 800 the government says operate today. And our estimates of when they went up are inexact. That means our tally of more than 1,050 people who have died near them is an undercount. And we don’t have a complete picture of where along the border the problem is worst.

An internal audit would not have such limitations. In theory, CBP tracks when each tower went up and when it was active. It has had the ability to audit their effectiveness—while quantifying how many people have died near them—since at least 2015, when it began installing the types of towers included in our analysis. But sources told us it has never done such an audit. Our findings make the need for one more urgent.

2. Fix the way the agency tracks migrant deaths and remains

If CBP is to analyze instances where people have died near its surveillance technology, it needs to know when and where those deaths happen. And the current system for tracking that data is beyond broken.

As we note in our methodology, there is no national agency that collects and publishes all available records on migrant deaths. We sifted through thousands of pages of records from dozens of agencies in Texas alone, and our picture is still incomplete owing to the lack of a shared protocol for documenting these deaths in public records. 

Border Patrol has had its own internal system, called the Border Safety Initiative Tracking System (BSITS), in place since 2007. But it wasn’t until 2021 that national rules were established for how different sectors should enter cases into the system. The next year, the Government Accountability Office (GAO) found that many cases still weren’t being tracked in this system, especially deaths that weren’t discovered by Border Patrol. In 2023, the GAO reported that Border Patrol has taken steps to fix that, but outside researchers consistently find the agency’s estimates to be undercounts even today.

In addition to tracking migrant deaths, CBP launched a program in 2017 to prevent them, called the Missing Migrant Program. In April 2025, the GAO found the program was improving the way it collected and tracked data about events like 911 calls and rescues. The problem, it said, was that there was no plan in place to use the data to measure whether the program was actually increasing rescues or reducing deaths, which are its primary goals. It made two recommendations for solutions, neither of which has been implemented.

3. Track cases when surveillance technology leads to an apprehension

In 2014, when the virtual wall was much smaller, the GAO suggested that Border Patrol agents log when a piece of technology helped them apprehend someone who was trying to enter the US illegally. It would take agents some extra time, but logging these assists would help Congress understand how often various surveillance technologies were leading to actual apprehensions. Once this data was collected, GAO added, the agency should measure whether its surveillance technologies were working. CBP agreed, and it started requiring agents to enter such information.

Then, in 2017, the GAO reported that the data being collected was inaccurate and unreliable; it found agents in Texas, for example, who were crediting a specific type of tower in apprehensions hundreds of times, even though no such towers existed anywhere in the state. Border Patrol again said it would improve; it agreed to train agents on the importance of collecting such information.

But years later, agents told us, it’s still not a priority. One former CBP official described the mandate as “spitting in the wind, to be honest,” and said the messy reality of apprehensions doesn’t translate into neat metrics. (The official, who requested anonymity, served during the Biden administration and wasn’t authorized to discuss sensitive issues.) 

And in August 2021, more than seven years after it agreed to do so, CBP still had not produced any analysis on the towers’ effectiveness (it finished one in 2022, the GAO notes, but the results are not public).

That means the virtual wall is growing faster than anyone is assessing its impact.

4. Investigate deaths as potential surveillance failures

CBP should also examine new deaths as they occur and determine whether they happened within sight of a tower or other surveillance technology. And for those that did, the agency should determine why those people weren’t spotted sooner.

Several agents and officials told us that if Border Patrol agents come across the body of someone presumed to have died while crossing into the US, there’s no requirement to check whether that person died near surveillance infrastructure and no process for doing so. At a minimum, that means potentially valuable evidence about the death goes unexamined. It also means that migrant deaths are not routinely treated as possible failures of the virtual wall.

CBP should automatically check every migrant death against its internal database of towers and investigate those that occurred within surveillance range to determine what the system detected and whether agents responded appropriately. Agents already log GPS coordinates when they apprehend a group or encounter footprints, among other events.

Without that check, law enforcement, medical investigators, and families may never learn that someone died near a surveillance tower—and that footage of the moments leading up to the death may exist. CBP typically deletes surveillance footage from the towers after 30 days unless an internal review is opened, which usually happens only for deaths in custody. At a minimum, flagging these cases could preserve footage for those investigating the death and give families a chance to request it.

Such reviews could also expose flaws in the virtual wall itself. Consider one location we identified in New Mexico, where several people died despite being in clear sight of an advanced AI tower. They died within the span of a few months, sometimes just hundreds of feet from one another. Investigating these cases might reveal, for example, that a camera repeatedly failed to detect or identify people in a particular area, or that agents were alerted and simply never responded.

Carrying out these postmortem inquiries would be the only way to answer the question that families, researchers, and volunteers raised most frequently in our investigation into the hundreds of migrants who died near towers: Did the cameras see them, and nobody came? Or did nobody see them at all?

When José Morales Bernal crossed the border into the United States on April 8, 2024, the day before his 32nd birthday, it should have triggered a chain of technological alerts and human responses. 

As he walked through the desert in southern New Mexico that morning, he was within range of three surveillance towers. Newly installed by US Customs and Border Protection (CBP), they were built by the defense tech company Anduril and equipped with cameras and AI to automatically detect and track people. They transmit live video to nearby control rooms and can send alerts to the government-issued smartphones held by agents in the area, prompting the closest available to respond.

These AI-enabled towers are meant to give greater visibility across the 1,951-mile southern border, freeing up border agents from having to spend hours staring into video monitors. They were installed in this particular place to spot border crossers before they reached the nearby town of Sunland Park.

If the system worked as intended, Morales should have been apprehended. If he needed medical help, agents were trained to provide it.

That didn’t happen. Despite the nearby surveillance towers, it was employees of the local landfill, rather than Border Patrol, who first spotted Morales that morning. At 1 p.m. the workers saw him again, now lying in the sand. At 4 p.m. the landfill workers saw that he had not moved and called Border Patrol. Agents arrived 45 minutes later. He was dead. When an agent then called 911 to report the body, he said it was “probably one of the migrants crossing through there,” seemingly unaware that Morales had been moving near the agency’s surveillance systems earlier that day. 

Morales had died just 360 feet from the closest surveillance tower. Two more towers stood watch to the east and the west. An autopsy later concluded that he had died of “environmental exposure.” 

The towers surrounding Morales were only the latest addition to the “virtual wall” the government has spent 25 years and billions of dollars building along the entire border, which also includes earlier generations of towers with more basic features, as well as blimps, drones, seismic sensors, and even tunnel-sensing robots. Together, all this technology provides “persistent surveillance” and “situational awareness” to help the Border Patrol quickly and accurately detect people crossing and, crucially, make sure agents are sent to intercept them. CBP has also credited it with saving lives.

But a first-of-its-kind investigation by MIT Technology Review reveals that deaths like Morales’s are startlingly common. We cross-referenced nearly 4,000 locations where human remains were found—drawn from records collected by nonprofit groups like No More Deaths and Humane Borders, as well as hundreds of records we obtained in Texas—with information on nearly 600 towers identified by the Electronic Frontier Foundation. We considered when towers were installed and when each person is estimated to have died, and combined this information with field reporting from the border. The result is the first comprehensive map and analysis of deaths near CBP surveillance towers. 

Our investigation shows that a humanitarian crisis at the border has unfolded in view of the government’s own cameras.

We found more than 1,050 people who died within range of border surveillance towers between 2015 and early 2026. These deaths are not failures of a few towers or technologies: We found deaths within the advertised range of nearly two-thirds of all the towers we analyzed. Our topographical analysis—which assessed the degree to which terrain might block a tower’s view of a particular death and its surveillance area in general—found that some have sight of as little as 10% of their advertised surveillance area, and yet we also found most deaths did not occur in towers’ blind spots. These deaths are not the result of legacy systems, as our estimate found more than 110 people have died within range of modern autonomous towers from Anduril, among the most advanced systems CBP has deployed, since 2021.

The overall picture reveals repeated failures of one of the virtual wall’s basic security functions, as CBP has described it in press releases: to effectively identify and locate migrants entering illegally into the United States. 

The year that Morales died, 18 other people died in that same stretch of desert in range of the three Anduril towers. Five were visible from the same surveillance tower closest to where his body was discovered.

One man, after walking a mile past the border and in range of two AI towers, dragged his 30-year-old brother into the shade when he began having trouble breathing, according to records we obtained from the medical examiner. The two were spotted by Border Patrol only when the man waved down a helicopter for help; by the time agents arrived, his brother was already dead. A 24-year-old woman died near another Anduril tower, where our terrain analysis showed it should have had clear sight of her location. Her body lay unnoticed for weeks; it was decomposed, and blistered by the summer heat, when it was spotted by agents patrolling the area. 

A few miles west, and a short drive from a Border Patrol station, four other AI towers stood watch. Near them, 10 more bodies were found that year, all in locations where at least one tower had a clear view.

In response to a list of questions, an Anduril spokesperson replied that once a tower is delivered, it is operated by CBP, and directed questions about specific incidents to the agency. The response noted that an incident occurring nearby does not mean the tower missed a detection and said actual surveillance ranges vary depending on terrain, physical obstructions, and the boundaries CBP sets for where the tower should look (CBP is able to set virtual boundaries on towers’ views for privacy and other reasons). The Anduril spokesperson also alleged inaccuracies in our reporting, given those boundaries and obstructions, but did not respond to follow-up questions on what was inaccurate.

The most pressing question in any death near the virtual wall is whether authorities knew someone was there and failed to reach them or weren’t aware anyone was crossing at all. Either is a system failure. And the deaths we found capture only the failures that left a trace—we don’t know how many people pass through undetected, or how many bodies remain undiscovered.

Interviews with more than 45 people—including current and former White House advisors and presidential appointees, Border Patrol agents, medical examiners, sheriffs, humanitarian volunteers, and employees of tech companies—showed that both types of failures are occurring: The technology is failing to detect, and agents are failing to respond. Both show the limits of throwing technology at a complex problem. 

The findings reveal previously unreported issues with the virtual wall, even as it continues to enjoy broad political support and a surge in federal spending. Border security hardliners have long seen it as another tool for stopping smugglers moving drugs or people, while others tout it as a cheaper alternative to a physical wall. In 2023, the government estimated that its plans for using the towers, which now number 803, would cost $6.2 billion over their lifespan. With the historic levels of funding it was awarded in 2025, CBP plans to spend $1 billion for 1,497 more towers by 2034.

But our reporting shows that CBP has done little to assess how its towers are working or how many people have died where they keep watch.

Officials who oversaw border security across the last four presidential administrations told us they believed deaths near the so-called virtual wall were either exceedingly rare or nonexistent. But our reporting shows that is not the case. None could point to any comparable analysis the government had ever conducted on its own. Former agents and officials also told us that when someone’s body is found, CBP does not formally investigate whether surveillance should have detected them or, if they were detected, why agents didn’t reach them before they died. 

In response to nearly 30 questions about our findings, which covered multiple generations of technology, Hilton Beckham, CBP’s assistant commissioner for public affairs, said, “Autonomous surveillance towers use artificial intelligence to detect and classify people, vehicles, and animals and alert Border Patrol agents to activity in monitored areas … ASTs complement physical barriers and other border security infrastructure by improving detection and situational awareness between ports of entry. CBP evaluates the technology based on its impact on detection, response coordination, agent safety, and mission outcomes.”

“I’m sure that the cameras and other surveillance assets do deliver … useful intelligence and enhance operational efficacy in some places, under some conditions, in certain circumstances,” says Geoff Boyce, an assistant professor of geography at University College Dublin who has studied surveillance technology used at the US southern border. “I’m also absolutely positive—because this has been the track record—it is not delivering the level of operational support, information, or efficacy that either the companies delivering these infrastructures or the Border Patrol and Department of Homeland Security claim.” 

We also reached out to multiple lawmakers from both parties with a summary of our findings. In response, Delia Ramirez, a Democrat representing Illinois’s 3rd district who sits on the House Homeland Security Committee, said, “AI-powered surveillance technologies are not making us safer. Yet DHS continues to spend millions of taxpayer dollars on these ineffective, negligent technologies, with no commitment to oversight or transparency … It is clear we must terminate CBP’s integrated surveillance tower program and dismantle DHS.”

Building the virtual wall

Since its earliest efforts to police the border with Mexico, the US government has faced the same basic challenge: How do you effectively secure nearly 2,000 miles of remote and rugged terrain? It has increasingly turned to technology for answers.

Gerardo Galvan joined Border Patrol in 1995, as the government was undertaking an unprecedented expansion of border enforcement. In 1993, President Clinton’s first year in office, the agency mobilized huge numbers of agents to guard the country’s urban borders, starting in El Paso, Texas. That pushed more crossings to rural, unpopulated areas—which quickly proved to be far more deadly because of the rugged desert terrain, extreme temperatures, scarcity of water and the long, indirect routes often used by smugglers.

Stopping these remote crossings was a new law enforcement challenge for the agency, especially given the limited technology available at the time. “We had radios,” says Galvan, who would go on to be the head of operations for the El Paso sector—not the cell phones or GPS systems agents have today. They relied on underground sensors left over from the Vietnam War to alert them to movement.

The biggest technology rollout of Galvan’s early career started around 1998. That was when Border Patrol began installing its first video cameras that agents could remotely pan and zoom. They were mounted on metal structures similar to cell-phone towers, 60 to 80 feet tall, or on buildings in urban areas. 

“For the first time, somebody sitting in a room somewhere was able to surveil large swaths of area without being dependent on getting an agent out there with binoculars,” says Matthew Hudak, a former deputy chief of Border Patrol who was working as a frontline agent at the time. “That was a very significant game changer, and for the most part, it was a huge, huge advance.”

This early iteration of the virtual wall was a big deal: In its first eight years, 200 towers were installed along the southern border at a cost of more than $429 million. But the system’s flaws were obvious to anyone who used it.

“There could be … 30 screens on a wall,” says Mark Borkowski, who managed CBP contracting, including the surveillance tower program. Each camera might be capable of seeing three or five miles in any direction. The agent was expected to keep eyes on all that space—upwards of 850 square miles.

“Well, what are the chances after 20 minutes that those Border Patrol agents would see an explosion on one of those screens?” Borkowski says. “About zero.” 

When these cameras were first being put up, the plan was to pair them eventually with a system that would automatically direct them toward places where sensors picked up activity, so agents would know where to look. But by 2005, the DHS inspector general said that still hadn’t materialized, and that illegal activity might “go unnoticed” unless agents were actively watching the cameras. Nonetheless, more towers were installed, forming what is today known as the Remote Video Surveillance System (RVSS). In 2013, the defense giant General Dynamics won a contract worth up to $103 million to overhaul and expand RVSS along the border, becoming the primary contractor behind the system that remains in use today. (That contract eventually grew to a ceiling of $216 million, and the company received another award in 2023 worth up to $135 million.)

General Dynamics would install more cameras, better sensors, and additional towers. But the problem identified nearly 15 years before remained unsolved. It was information overload: One agent told MIT Technology Review about an incident in San Diego in which cameras captured a group of 30 crossing the border and getting picked up in a van, all unnoticed by the agent monitoring those cameras. Agents said the virtual wall was bringing more of the border into view than they could reasonably pay attention or respond to. 

The camera room was also the job nobody wanted. New Border Patrol agents, often in their early 20s, want to be outside in trucks and ATVs, not sitting in a dark room watching screens—even though the expanding virtual wall increasingly required someone to do exactly that. “Someone that’s injured and can’t go out into the field—we’ll send them to the camera room,” says Rafael Reyes, who was in charge of the Border Patrol station in Deming, New Mexico, until 2024. 

Reyes says it was also a job his agents were often too busy to do properly. His station area had about 30 cameras, and the agents responsible for them also had to monitor sensor alerts and radio traffic, run records checks, and coordinate with other agents. “Then if they have some time, they’ll monitor the cameras,” he says. 

Not only that, but many cameras didn’t work. At Reyes’s station in New Mexico, five of the 30 cameras were broken on any given day, he says. Others had limited functionality; they might be capable of looking in just one direction or seeing only in the daytime. Jaime Fierro, a former agent who worked in the Laredo sector in Texas until 2025, said the broken towers were tanking morale among agents and that they brought it up at every staff meeting. “That was like basically having an eye shut for us on the border,” he says. 

They weren’t alone: In 2024, members of the House Committee on Homeland Security wrote that more than 66% of Border Patrol’s first-generation towers were unusable, despite operating budgets of $50 million to $100 million each year. In January 2026, 30% remained broken, one congressional staffer told MIT Technology Review. The RVSS program was estimated in 2020 to have a total cost of $3.7 billion.

Border Patrol has nonetheless publicly credited these cameras with helping to save lives. The agency has published dozens of press releases, dating as far back as 2016, about instances when agents responded to people seen in distress on the remote video cameras. Despite these publicized wins, the bodies began to pile up, undetected, beneath these systems’ watchful gaze.

Near El Cenizo, Texas, there’s a roughly 13-square-mile expanse of mesquite bushes and grasses described by the Webb County Sheriff’s Office as containing sections of Hachar Ranch and Espejo Ranch. It’s surrounded by seven RVSS systems. It was nonetheless a hot spot for fatalities: MIT Technology Review found that 19 people died there, including one man who died less than 500 feet from the road bordering the ranches and half a mile from two schools. In each of these cases, narratives that we obtained via public records requests describe officers’ learning of these people only after their bodies were discovered or through 911 calls reporting them missing, not as a result of the border surveillance technology installed across Hachar Ranch. 

In response to a list of questions about its towers and these specific incidents, General Dynamics referred us to CBP. CBP did not address any of our questions about RVSS towers.  

 

Stories of some of the people that died near RVSS towers

Many deaths near RVSS towers happened in Texas, in the heavily surveilled corridors between the Rio Grande and the nearby highways, where paid smugglers often organize pickups after crossings. Cameras watch many of the spots where people are known to cross, often looking across the river into Mexico to spot them before they reach the water. Drownings are common; we found multiple cases in which local fishermen hooked the bodies of people who had died attempting the crossing.

In one instance, a family of three from Brazil attempted to cross the Rio Grande near Del Rio, Texas, in 2022. The father, Daniel Lenda, was carrying his two-year-old daughter, Eloah, on his shoulder when he fell in the water. The mother, 23-year-old Thais Natali Montenegro Lenda, lifted Eloah out of the water as she watched her husband disappear. 

She placed the girl, now unresponsive, on a rock and went for help. She was just 100 feet from one camera system and a third of a mile from two others, set up precisely to enable faster apprehensions at a popular crossing point. Our topographical analysis shows that all three cameras had a clear view of the place where the family crossed and exited the water. Nonetheless, Thais Natali was not spotted until she waved down an agent on the Del Rio-Acuña bridge. The agent provided CPR to the toddler, who did not survive. Daniel’s body was found days later. 

Val Verde County sheriff Joe Frank Martínez responded to that case (though Border Patrol might discover bodies, local law enforcement is responsible for investigating the deaths). He grew up in the area with his nine siblings and has been sheriff since 2009. Riding with MIT Technology Review through Del Rio in his patrol vehicle, he rattled off the deaths that his office has responded to since he became sheriff. 

He keeps a three-inch-thick binder of these cases in his office. But this case with the toddler still sticks out in his memory. He told MIT Technology Review that when he arrived on scene, he heard from Border Patrol that the tower operator monitoring the camera had seen the father and daughter fall in the river. He never received an answer as to why agents weren’t sent sooner. Martínez asked for the footage from the surveillance cameras to aid in his investigation of what happened. Border Patrol never provided it. (General Dynamics did not respond to our questions about this incident, instead directing us to CBP. The agency did not address questions about this incident either.)

Establishing how many people have died in range of the virtual wall is challenging, but it’s especially so in the case of the RVSS towers. There are different models, some with a shorter range of sight—one to three miles—and others with a longer range of up to 7.5 miles. Former agents and officials told us most towers should see at least five miles, but since no public records exist to confirm the model type of any given tower, we used several possible viewing distances to calculate the number of deaths considered in range.

Another complication is that many remains are never found at all. For those that are, MIT Technology Review could analyze only deaths that had been logged with GPS coordinates or sufficiently precise location descriptions. Many records lack this level of detail, including nearly a fifth of the more than 1,500 cases we obtained from 14 Texas counties. We also needed to determine when someone died, rather than simply when their remains were found, to confirm that a tower was present at the time of death. When official sources did not provide an estimate, we consulted multiple medical examiners on how to make that determination from the available details. And satellite imagery provides only intermittent evidence to suggest when towers arose, not exact proof of when they were functioning. For these reasons, as detailed in our methodology, we tried to be conservative in our analysis.

Even so, the number of deaths estimated to have occurred near at least one of these towers is staggering: somewhere between 500 and 700 since 2015. Even if we look only at towers in rural areas, where their views are far less likely to be blocked by buildings, we still estimate more than 300 people have died in places where the cameras had a line of sight, most less than three miles from a tower. Of the nearly 300 RVSS towers we included in our analysis, more than 250 were within range of a location where someone died. More than 150 people died within a half-mile of one, often across open, empty plains, close enough to see the cameras themselves perched atop towers up to 200 feet tall.

The next watchers

When Borkowski was overseeing the rollout of those first-generation towers at CBP, he became increasingly certain that simply having more remote video cameras was not going to lead to faster responses. The information overload meant the government could watch a large area of the border but not meaningfully see, much less act on, everything within it. The agency’s focus then became solving that overload problem. 

What if, instead of relying on someone to scan 30 screens, the towers could use radar to detect motion, and then point the camera to it automatically? 

In 2011, Border Patrol asked the industry for a new generation of technology that could do just that—initially focusing on Arizona, where the border was busiest. The contract ultimately went to the Israeli defense contractor Elbit, whose tower systems watch Israel’s borders. In 2015, the company started installing versions of these—called integrated fixed towers, or IFTs—that could each see more than five miles.

The IFTs were still being installed when President Trump first took office, in 2017. They didn’t command the same attention as Trump’s campaign promise of a physical wall, but there was still support. Trump’s deputy CBP commissioner, Ronald Vitiello, told Congress in 2018 that the IFT towers automatically detect people with radar and then track them, adding that such surveillance technology is “critical in protecting border areas with short vanishing times, where illicit crossers can quickly evade law enforcement by ‘vanishing’ into border communities.”

There were 55 IFTs built in Arizona under an initial $145 million contract with Elbit and other contracts that followed. The towers collectively watch more than 7,000 square miles of terrain, from the mountainous areas outside Nogales to the plains of the Tohono O’odham Nation Reservation. Arizona is in many ways well suited for these tall towers because much of its landscape is flatter than, say, Southern California. 

But even here, being within a tower’s advertised range does not necessarily mean being within view. There could be terrain, brush, or buildings that block the line of sight. 

MIT Technology Review conducted a topographical analysis to estimate whether each tower had a clear or obstructed view of the place where someone’s remains were found. We estimated the height of each tower and used US elevation maps to model the terrain between it and the remains. The analysis does not account for buildings or vegetation, though all the towers we analyzed use some form of thermal imaging that is designed to see through light brush.

Most of the instances in which someone died in range of a tower happened where we estimate that the camera’s views were not blocked by terrain—a finding that held in both rural and urban areas. And even when someone died in a blind spot, it does not mean they did not pass through a tower’s line of sight. Additionally, these numerous blind spots indicate issues with the tower’s supposedly comprehensive coverage.

We found that nearly 350 people have died in range of Elbit’s towers since 2015. There were 40 people who died within range of four or more IFTs, with some near as many as six different towers. In May 2024, a 19-year-old man died where three IFTs had a clear view, and a year later, in May 2025, a 38-year-old woman died where two IFTs had a similarly clear view. Elbit did not respond to a detailed list of questions about its towers or deaths we found to have occurred near them. 

There’s been a death within range of 52 out of the 53 IFTs we analyzed, but some were particularly notable. There’s a tower located a couple of hours from Tucson on the Tohono O’odham reservation that we estimate has upwards of 80% visibility of the surrounding area. Still, 27 people, ages 19 to 64, have died in its surveillance area just since 2021, along with several others whose remains have not been identified. In 25 of these cases, we estimated that the tower had a clear line of sight to the location where the person died.

Stories of some of the people who died near IFTs

---

The age of autonomy

Around 2018, officials within Border Patrol were again discussing the limits of the towers. The new radar systems were better, but they could only detect movement, not what was moving. This generated false alerts from cattle, tumbleweeds, and other harmless activity. The now decades-old dream of the virtual wall remained unfulfilled. Under pressure from Congress, which was demanding to know what $33 billion in proposed funding for border security over the next decade was going to achieve, Border Patrol once again was hoping a new generation of technology could solve its problems.

That year the head of CBP, Kevin McAleenan, created a unit called the Innovation Team. “He gave them some seed money to go after kind of Silicon Valley innovative technologies and run pilots,” Borkowski says. 

One company they landed on was Anduril, the defense tech startup founded by Oculus VR founder Palmer Luckey. The company was without a product but had funding from Peter Thiel, recruits from the security tech giant Palantir, and a goal of bringing Silicon Valley experimentation to defense and national security technology. 

With grants from the Small Business Innovation Research program, Anduril’s product started taking shape: a 33-foot-tall solar-powered structure with cameras to see and radar to detect movement, all sitting atop a modular metal pole, with solar panels arrayed beneath. 

More important was the component you could not see: computer vision that promised to automatically classify, track, and generate alerts for what came within the system’s view. Anduril’s standard tower couldn’t see as far as the Elbit and General Dynamics models—about 1.75 miles instead of more than five—but, the company argued, it didn’t need to: By automatically identifying and tracking people, the towers could give agents enough warning to respond. (Its newer, extended-range towers, on the other hand, stand 80 feet tall and are advertised as able to detect and track “objects of interest” up to 7.5 miles away.) 

When President Trump left office in 2021, Anduril continued to find support from his successor. Joe Biden had campaigned on the promise of shutting down the project to expand the physical wall, but he saw political reasons to keep investment in tower technology on the table. “It’s less visible—it’s unobtrusive,” a former immigration advisor to Biden recalls, adding that despite some privacy concerns, the towers were mostly uncontroversial. 

When Border Patrol deployed Anduril towers to the Santa Teresa station in New Mexico in 2021, a press release stated, “Because of its accuracy with detection, in many cases this type of technology can and will save migrant lives.” In a permitting request filed in 2024 for towers in California, the government was even more explicit: “If anyone within the viewshed of the towers appears to be in distress, EMTs or first responders will be sent to help.” 

Anduril echoed those claims. The software built into the surveillance towers, the company wrote in a 2021 press release, gets agents “out of communications operations centers and into the field where they can effectuate security and humanitarian responses.”

It wasn’t the first time CBP had discussed a humanitarian side to its duties; in 1998 the agency created BORSTAR, a search and rescue unit, and by August 2023, it had installed some 170 rescue beacons across the border that people attempting a crossing could theoretically use to call for help. But they weren’t effective, says Mario Agundez, a retired Border Patrol supervisor from Arizona who had worked on various rescue efforts; people either avoid them or “die next to them, because there was no way [for Border Patrol] to get to them in time.” (They don’t always work, either; one that MIT Technology Review passed, just next to a dirt road in the Otay Mountain wilderness area in Southern California, had its prominent solar-powered blue light, meant to be visible for miles to people in distress, turned off on a recent night in August. A CBP spokesperson did not comment on why it was off but said, “The solar-powered beacons are maintained and monitored by the responsible Border Patrol station that patrols that area.”)  

Anduril soared through a pilot phase in just two years before its Sentry towers—then the only autonomous surveillance towers CBP was using—moved into an official program of record with the government in 2020. That’s light speed in the world of government procurement, especially since it was Anduril’s first hardware product. By June 2021, CBP had already spent over $98 million (Anduril’s contract for these towers is now worth up to $1.1 billion). 

---

Gerardo Galvan was in charge of the Santa Teresa Border Patrol station—which patrols the area in New Mexico where Morales’s body was later found—when it received its first batch of Anduril surveillance towers in 2021. His agents were worn out; the El Paso sector, home to his station, had seen encounters tick up each month since April 2020. Galvan says the agents were surprised but eager to get the AI towers, and he began planning where to place them. The sites he was considering sat where one kind of landscape abruptly gives way to another. 

To the east of the station is Cristo Rey, a mountain that obscures the view of the city of El Paso. Running north from Cristo Rey is the state’s border with Texas, where the banks of the Rio Grande are flanked on either side by pecan orchards. And west of those orchards are the towns of Sunland Park and Santa Teresa. The last houses in these neighborhoods butt right up against the desert, which sprawls westward without a single gas station for nearly 100 miles.

It is in this stretch—between the US-Mexico border to the south and Highway 9 to the north—where agents from his station aimed to stop migrants. And when Galvan was planning the tower locations, there were lots of crossings; Santa Teresa was becoming the busiest station in the 125,500-square-mile El Paso sector, as heightened border enforcement in Texas drove people west. Along this new route, more people were dying than in previous years. (The traces of that period are still visible years later; MIT Technology Review visited the area with the humanitarian group Battalion Search and Rescue and saw jawbones bleached white by the sun, alongside frayed backpacks and clothes.)

Galvan’s aim was to use the new towers as a stopgap where there was little other border infrastructure. That included a spot near the foot of Mount Cristo Rey, where there was a break in the border fence and people often attempted to cross. Galvan looked at data on previous apprehensions with an eye toward helping agents spot people before they arrived in more residential neighborhoods. Border Patrol struck agreements with private landowners, who generally didn’t mind having Anduril’s low-profile towers on their land (and were compensated via lease agreements). 

When the towers were first set up near the Santa Teresa station in 2021, engineers from Anduril came to fine-tune the algorithms meant to autonomously classify whether what it had detected was a vehicle or a possible border crosser, among other things. A former engineer for Anduril, who spoke on the condition of anonymity to discuss his previous employer, says these algorithms were the main focus of the tower program; Anduril didn’t manufacture the cameras or radar systems itself, so the algorithms were its main value proposition to Border Patrol. 

Galvan says the towers would flag people crossing before agents saw them. But he also saw problems from the start. Agents would apprehend a large group near one of the towers, for example, and check to see if the footage revealed anyone who got away, only to find that the incident hadn’t been captured at all. Anduril’s towers constantly pan around, lingering only on an object of interest. But in these cases, Galvan says, they panned elsewhere, failing to capture the border crossers they were supposed to automatically track.  

Reyes says algorithm mistakes were infrequent but problematic at his station. A group of people was sometimes labeled as cattle, for example, and agents wouldn’t receive an alert. The problem got worse the farther groups were from the camera.

Anduril has touted its technology’s ability to reduce false positives, like cases in which wildlife is mistaken for people. But the company has said little about these false negatives: people the systems fail to detect.

Jaime Fierro, the former agent in the Laredo sector in Texas, describes one incident that shocked him: Agents pursued a car believed to be carrying migrants who had just crossed the border. The car turned around, drove toward the banks of the Rio Grande less than 100 feet from an Anduril tower, and crashed right into the river. Several occupants got out and swam to Mexico while the car floated in the water. Back at the station, Fierro hurried to see what the camera had picked up. Only it never detected the incident at all. There was no footage. “That was a huge, huge issue,” Fierro says. He remembers Anduril coming out to investigate and the problem being sent all the way up to Washington leadership. But Fierro never got an explanation of why the crash was missed. (An Anduril spokesperson did not respond to a question about this incident but said the range of the towers we asked about was limited by physical obstructions and by boundaries established by CBP. )

Several agents estimate that the towers initially missed 10% to 15% of what they should in theory have caught. “To us, 10-15% on a system we spent millions on that was supposed to work—that was alarming,” Fierro says.

Philip Sullivan worked as an agent in the Laredo sector too. When the Anduril program started, he says, he put his hand up to be involved and even went to an Anduril test site for training. When new towers went up, he worked directly with the company on improving its algorithms. He enjoyed the work. But he describes one recurring issue they couldn’t resolve.

Smugglers would put up to a dozen people on a rubber raft, cover them with a tarp, and cross the Rio Grande while using submersible motors to propel it forward. The tarp fooled the algorithm, which detected the motion but attributed it only to an “unidentified object.” That meant no audible alert—the very feature agents relied on, because the Anduril system was supposed to do the watching for them. 

The smugglers repeated the tactic dozens of times. Other smugglers would cover groups with netting or blankets so they would remain undetected while walking through the brush. Only afterwards, having seen the group farther inland and noticed clues about where the people had crossed, would Sullivan review the footage and see the crossings that were missed. He worked with Anduril, sending engineers videos of the crossings to retrain the firm’s algorithms. But the situation had not improved by the time he got promoted to a sector-level job sometime in 2022, he says.

“The whole purpose of their system,” Sullivan says, was the idea that “the cameras are moving around, identifying automatically and tracking and recording and flagging.” Having cameras that could pan on their own worked wonders, he says. But the algorithm had its limits. “You look at it on the screen yourself, and you see the blob and the raft moving across—well, you know that’s people. But to get an AI to identify that is a challenge.”

When the towers were first being rolled out at the Santa Teresa station in 2021 and 2022, Galvan says, similar issues led to a negative feedback loop: An agent would see a tower miss something significant and come to distrust the algorithm. When that agent’s turn came to operate the towers and manage the alerts, they might override the AI altogether and use the camera manually. That would lead to more missed alerts, more distrust.

Despite these shortcomings, the El Paso sector’s relatively flat landscape meant that the system’s cameras had about 80% to 90% visibility, according to our topographical analysis. That was better than in hillier or more mountainous areas, like the Otay Mountain Wilderness to the east of San Diego, where we estimate Anduril’s towers could see just 15% to 25% of their promised coverage area. 

It was in this wilderness, in fact, that some of the first bodies started appearing near Anduril’s towers. During one week in August 2021, two women in their 20s, from the same city in Mexico, died miles apart. Their bodies were found near three different towers that were first observed on satellite imagery between March and July that year. 

One of them, 26-year-old Sarahi Hernández Alfonso, began her journey to cross into the US on August 5, and medical investigators note that she reportedly fainted the next day near the Otay Mountain Wilderness and was left behind. It would be another four days before the Mexican government reported her missing to Border Patrol on August 10, supplying a set of coordinates. Agents went to those coordinates and found her decomposing body. She’d died 1.3 miles from two different Anduril towers, one of which our analysis found had a clear line of sight.

Just the day before, 25-year-old Karina López Antonio was found dead near a third Anduril tower. She had crossed the previous night with her cousin and nephew. That morning, she felt sick, and her nephew looked for a Border Patrol agent to help. When agents returned, she was dead. 

The following October in New Mexico, where the new Anduril towers had been rolled out, agents were patrolling on ATVs when they came across footprints. They followed them until they found the body of 31-year-old César Perea Itzincab. The presence of maggots and the level of decomposition indicated that he had died at least a week earlier, and medical examiners on the scene believe he had dragged himself to the spot where his remains were found. He was 1.5 miles away from an autonomous tower—within Anduril’s advertised range—and should have been visible to the camera, according to our topographical analysis.

We can’t know what those cameras saw as these people died. Footage and data from Anduril’s towers are overwritten every 30 days, and a former official with internal affairs at CBP told us the information wouldn’t be saved for longer unless it was part of an active investigation—which starts only if someone dies in custody. CBP did not respond to questions about specific incidents like this one, and the Sunland Park Police Department closes cases like this when it determines that no crime has been committed. 

The lack of documentation is a missed opportunity, says Amerika Garcia Grewal, co-director of the Frontera Federation, which aims to help rescue migrants in distress—or find and identify their remains. “The surveillance towers along the border could be incredibly useful,” she told us. For example, they could in principle be used to hold Border Patrol agents accountable for whatever actions they did or did not take to locate someone in distress. “What the tool turns out [to do] depends on the person who is holding it,” she says, adding, “I don’t trust the folks that are using them.”

“If I could get that footage,” she says, “then we would go through that, and hopefully bring some answers” to family members and “documented proof of what happened” in their loved ones’ final moments. 

When asked about deaths near Anduril towers, agents often point to faults in the technology. Galvan, for example, says the towers don’t capture how groups move: A large group of 20 people might splinter into smaller groups, especially if agents are pursuing them, and the tower doesn’t keep track of everyone. If someone’s missed, he says, “that person stays behind in the brush, out of sight.” Then they might pass out—which could be a death sentence in the desert. 

And though Border Patrol says the towers can “hand off” surveillance of people from one tower to another, agents on the ground say that people are often missed during these handoffs. On top of that, the algorithms remain imperfect, according to Jason Owens, the chief of Border Patrol from June 2023 to March 2025. “We never really got to the point of ‘Set it and forget it,’” he says.

Agents also say they were just stretched too thin to respond effectively. Around 2022 and 2023, the border saw historically high levels of migration. Title 42, a policy that began under Trump and continued under Biden until May 2023, cited a public health emergency to expel migrants before they could apply for asylum. But the rapid expulsions carried fewer of the repercussions that could normally follow an apprehension, and many people simply tried to cross again.

Meanwhile, asylum claims had been rising for years, while shifts in US immigration policy and in the places migrants where originating from meant more people required lengthy processing rather than being quickly returned across the border or to their home countries. That increasingly tied up agents with people who had already been apprehended, they said, leaving fewer available to respond to surveillance alerts. 

In other words, from the agents’ perspective, the effectiveness of Anduril’s towers was limited by the same issue that had always caused trouble: There just weren’t enough agents.

But the overwhelming demands on agents cannot explain all the deaths our investigation found. In 2024, the border started to get quiet again. By July of that year, the number of Border Patrol encounters nationwide had plummeted from their highs in December 2022. Encounters in the El Paso sector, where the Anduril towers in New Mexico were located, had fallen to nearly one-tenth of their peak. Agents were less tied up than they’d been in years. The technology was supposedly improving, too. 

Galvan had left for a promotion by this time, but he says agents had come to trust the towers more, and they helped train the algorithm by giving a thumbs up or thumbs down if the system identified something correctly or incorrectly. There was also a new smartphone program: Every agent at the station was given an Android device with a map of where other agents were, and it sent alerts from the station about what an Anduril tower detected. 

Despite all this, the deaths near towers continued. There was Morales in April, whose body Border Patrol learned about only from the landfill workers, though it had lain for hours within sight of an Anduril tower. 

Others were discovered by luck. 

On June 20, Border Patrol agents were mistakenly tracking a group of people they thought might be migrants, though they were in fact employees of the same landfill. Those employees told the agents they had come across a body, which turned out to be partially mummified remains of a man in his 40s who died within range of two Anduril towers. 

Three days later, on the 23rd, another: the body of a 21-year-old Guatemalan man discovered by agents while on patrol, 1.3 miles from a tower across clear and open desert. Two days after that, another: An agent was looking for a lost person when he came across the remains of a 56-year-old man—just a football field’s distance from the previous one, with a similarly clear and unobstructed view to a tower.

So if Santa Teresa had gotten quieter—and become something of a showcase for Border Patrol’s newest technology efforts—what was going on?

“Agents review the information and determine the appropriate response—the technology does not make law enforcement decisions,” Beckham, CBP’s assistant commissioner, said. Meanwhile, individual agents described scenarios where they might not prioritize responding to alerts.

Sometimes “hanging back” was a strategy: Agents might observe a group to see what route they would take, or agents might not respond to a small group, in case those people had been sent by smugglers to distract from a larger group coming behind them. Other times, agents might deem the location where a group of migrants had first been spotted to be impractical for an apprehension, and instead wait to intercept them at another location.

Knowing someone’s location didn’t always lead to immediate action. Volunteers describe providing Border Patrol with the coordinates of someone who was lost but alive and then waiting, “sometimes [for] a week,” for agents to reach the location, says Garcia Grewal, from Eagle Pass, Texas. They might be told “Oh, so-and-so went out there and they found remains,” she recalls. “Well, they weren’t remains when we called you. They were alive.”

Mireya Morales, the younger sister of José Morales Bernal, who died by the landfill on the day before his birthday in April 2024, wasn’t aware her brother died near several surveillance towers, or that the government said these towers could save lives. “Well that’s good,” she said of that promise, “but in this case, it didn’t help my brother. And there’s no way to know exactly how things played out.”

“Border Patrol should have found him quickly,” she says, “and they could have done something.” If they did, this journey into the United States—his fourth trip—likely would have ended with his apprehension, detention, and deportation to Mexico. But he would have seen the birthday texts that his sister sent the next day. At least he would have made it home to celebrate his eldest daughter’s quinceañera, which took place earlier this year.

Stories like this are why Iván Chaar López, an assistant professor of American studies at the University of Texas at Austin who leads its Border Tech Lab, says he’d “rather talk about harms” than about “the failure of an algorithm.” He adds, “A system may fail, but humans suffer.” 

Stories of some of the people that died near Anduril’s autonomous surveillance towers

What happens next

Ultimately, any attempt to understand where things are going wrong is hampered by an institutional reluctance to measure the problem.

A complete audit of this virtual wall can only come from Border Patrol. The agency is, in some ways, increasingly equipped to take on that task: Agents told us that each time they apprehend someone, the location is logged with GPS coordinates, as are instances when Border Patrol sees evidence that someone crossed but cannot find them. It also has precise data on when each surveillance tower went up, which alerts came in, and how they were resolved. Having that information is the only way to measure whether the newest towers are truly missing fewer people than the previous generations. 

But researchers and outside oversight agencies say this ocean of data hasn’t translated into a reliable system for measuring the effectiveness of the virtual wall. For example, CBP records the people it misses by logging “gotaways,” a tally of how many times agents see signs of someone who evaded apprehension—footprints, appearances on camera, reports from other people who were apprehended. It’s imperfect. And that creates room for interpretation.

“The goals with border enforcement have always been a moving target,” says Jeremy Slack, a researcher of migration and border issues at the University of Texas at El Paso. “They’re always set up so that it’s a win-win.” If apprehensions go up, for example, CBP says it means agents have gotten more effective at catching people. If apprehensions go down, it means the border is quiet because people are too scared to cross. 

That’s not to mention the conflicting ideas within the agency about what the virtual wall was supposed to accomplish. Borkowski says higher-ups would ask how many fewer agents they could get by with if they built more towers. But supervisors receiving new towers told us they’d often ask for more agents, not fewer, because they now had more activity to respond to that had previously gone unseen. 

And if the goal was for the sight of the towers to deter people from crossing to begin with, an external study from RAND in 2020 was ambiguous: It found that deploying IFTs resulted in lower apprehension levels nearby but said this didn’t mean the towers were actually deterring crossings. (Research by Boyce and colleagues found that earlier surveillance towers in southern Arizona pushed people to seek more difficult terrain out of view.) It leads to a question: Do the deaths near the virtual wall constitute unacceptable surveillance failures, or are they within the range of effectiveness the government deems acceptable? (CBP’s response did not address our questions on how it explains these deaths.)

Against this backdrop, the Government Accountability Office and DHS’s Office of the Inspector General have tried to focus on a narrower question: Is the virtual wall increasing the likelihood of apprehensions? 

The answer has been incomplete since 2014. That’s when the GAO first suggested that whenever agents log their activity in Border Patrol’s database, they should include information about whether or not a piece of technology helped in their apprehension. This could at least do something to show Congress whether the technology is working. Border Patrol began collecting that information, but years of incremental improvements have been followed by repeated findings that the resulting data is unreliable or insufficient to determine how well the technology works. The reality, several people from Border Patrol told MIT Technology Review, is that agents see it as a chore demanded only by bureaucrats who don’t understand the realities of the border. 

It’s “spitting in the wind, to be honest,” a former high-ranking official at DHS during the Biden administration told MIT Technology Review on background. “If we have 10,000 people a day crossing in between the ports of entry, do you think an agent is worried about telling them how they freaking apprehended those people?” Several other officials said that the logging of technology assists had improved, but not consistently enough to help much with analysis. 

And there is no procedure—at the local or national level—to examine individual deaths near the towers or analyze them collectively for surveillance failures. “We never thought about plotting [migrant deaths] to see if the towers were missing things or agents were missing people,” said a former CBP official who evaluated how personnel handled deaths in custody.

“Nobody ever raised the issue,” the official said, adding that doing so could reveal important gaps in the agency’s approach: “If I still worked for CBP, and you brought it up, I would probably send some people to look.”

In response to questions from MIT Technology Review, CBP said that autonomous surveillance towers complement physical barriers by improving detection and situational awareness. Its statement did not address questions about its other technologies, or broader criticisms about how it evaluates the virtual wall.

This lack of measurement has not slowed the enthusiasm for more technology funding. The government spending bill passed in July 2025 awarded $2.7 billion for border security technology alone, and CBP was quick to signal to industry that the money would soon start flowing. “We’ve got an historic investment in infrastructure and technology coming up,” a CBP director told tech companies in an industry webinar that month. In a December 2025 interview, when he was director of homeland security and immigration at the America First Policy Institute, Cooper Smith described the technology investments as “fortifying” the border against a future president who might be, in his estimation, weaker on the border than Trump. (According to LegiStorm, a research organization that focuses on political staffers, Smith now serves in a policy-focused role with CBP.) The number of Border Patrol agents has climbed too, to nearly 21,500 agents as of June 2026—the highest in the agency’s history. 

In the Big Bend region of Texas, surveillance towers are becoming a focus of local politics: Officials and residents across party lines, including the sheriff of Terrell County, are asking for Anduril towers instead of a controversial new border barrier project that’s recently been halted.

CBP has announced a desire to spend $1 billion on nearly 1,500 more towers by 2034. The law now requires all these towers to be equipped with AI. In December 2025, Anduril’s Palmer Luckey said the towers remained the company’s best-selling product, one that gives “basically perfect situational awareness of what’s going on in the area.” Anduril’s towers have been sold abroad—to enforce the UK’s borders, defend US Marine Corps bases in Japan and elsewhere, and serve as drone defenses for an Australian Air Force base. 

But as for CBP, Anduril will no longer be the only game in town; General Dynamics has now released AI towers of its own and received an order for them in June from CBP worth up to $115 million. The Electronic Frontier Foundation—which has tracked surveillance towers since 2022—says some have already been spotted at the border.

This time around, the agency will be spending this money with even less oversight than it had just a couple of years ago. In October 2025, the Department of Homeland Security, which oversees CBP, dissolved its department-level office that oversaw major spending programs.

---

Most of the records on human remains that MIT Technology Review collected were current only through the fall of 2025 or, in a handful of cases, early 2026. Even those dates come with a caveat: Many remains are never found, and others are found months after someone died. The reports can then take months to be released via public record requests. That makes it nearly impossible to keep track of how many people have died near surveillance towers in anything close to real time.

Still, the border today is significantly quieter. The latest figures from DHS show just 128,009 “enforcement encounters” from January to August 2026, compared with more than 2.4 million during the same period in 2024 (“encounters” count each of Border Patrol’s interactions, not individual people).

Even as border crossing rates decrease and DHS speeds ahead with its tower acquisition plans, people are still dying within range of the surveillance towers.   

On September 14 2025, a 30-year-old Mexican woman named Graciela Gómez Hernández died just over 350 yards away from the physical border wall that separated Southern California from Tijuana—and within one mile of an RVSS tower. Earlier that afternoon, she had told her family that she could not walk any further, and her voice notes abruptly stopped. 

Her skeletonized remains were recovered nearly three weeks later, on October 4. Some of her bones were missing, and there was “apparent animal activity … to the ribs,” as the medical examiner’s report read. 

Sara Stroud, an organizer with the Borderlands Relief Collective, a volunteer humanitarian group that leaves water and supplies in a wilderness area frequented by people crossing the border, went to the site of her death a few days later to put up a memorial cross. A Border Patrol helicopter showed up almost immediately, flying low circles above the volunteers. 

Between the helicopter and the RVSS tower that was visible in the background, “I was astounded about how we were being surveilled the whole time,” Stroud says, adding that this was especially stark in contrast to the agency’s often slow—or absent—responses to people who’d died on camera. 

“Are you telling me you didn’t see it or you did see it?” she asks. “Because if you did see it, that’s horrible. And if they can’t, then how do you explain that?”

Additional reporting by Lillian Perlmutter.

This work was supported by a grant from the Tarbell Center for AI Journalism.

She had only walked for a couple of hours, and already she was lost. 

It was early afternoon on Sept. 14, 2025, when 30-year-old Graciela Gómez Hernández crossed the border from the eastern edge of Tijuana into Southern California, sending voice messages to her mother and sister as she walked. 

Gómez Hernández did not tell her family about her plan until the day before. Once she arrived in California, she hoped to work and send money back to her three children and her parents in southern Mexico. Her oldest son, age 11, had asked for a bicycle.  

She either climbed a ladder over the border wall or slipped through a section in steep terrain where fencing has yet to be built. Dozens of game trails cut through the dust on hillsides of sagebrush and manzanita. It was the first segment of a hike through the Otay Mountain Wilderness. 

The eastern edge of San Diego is just a mile away, but countless migrants hike 10 or 20 miles north through the mountains to evade the Border Patrol, emerging in smaller towns and on distant highways with fewer checkpoints.

Between 2 and 3 p.m., her sister said, Gómez Hernández sent a final message to her mother. “I can’t do it anymore,” she whispered into her phone. By the time a forensic crew came to collect her, two weeks later, most of her body had disappeared into the earth. 

She died on her own, but she was not, strictly speaking, alone. She was in one of the most heavily surveilled strips of land in the world.

Before she died, Gómez Hernández told her family in a WhatsApp message that she could see helicopters overhead and white Border Patrol trucks making periodic laps on the road below. And even if none of those agents saw her or knew she was in distress, something else had a clear view. Supervising the whole territory, from a nearby hilltop, was a slender gray surveillance tower. 

That tower, made by General Dynamics and in place since 2019, boasts two sets of electro-optical and infrared cameras with long-range lenses that can track targets 5 to 7 miles away. Gómez Hernández was just one mile away. 

Surveillance towers have proliferated in recent years, as contractors added newer camera systems that use artificial intelligence to distinguish border-crossers from other people and animals, and track their movements, at an overall cost of more than a billion dollars. 

Yet as the cameras scan the terrain, migrants continue to die in plain sight. 

Thousands of migrants have died while crossing the border in the past two decades, many of them near to Border Patrol surveillance installations. But the phenomenon is especially pronounced in the Otay Mountain Wilderness, just outside San Diego. 

A Times of San Diego analysis counted hundreds of cases along the California border—including at least 138 in the past four years—where migrants’ remains were found within the range of the very cameras designed to help track and intercept them. 

A topographical analysis conducted as part of an MIT Technology Review investigation estimated that more than half of these remains were found in a spot where a camera would have a clear view of a person, without terrain blocking the line of sight. 

Some of these migrants died swiftly, falling from the border wall or drowning in a canal or river. But dozens suffered slow deaths of exposure in the wilderness, deaths that might have been averted with emergency aid. 

People who died in Southern California came from just across the border, or from as far away as Africa. They were as old as 70 and as young as 5. They died in the heat or the cold. But they all died within range of a surveillance camera.

---

Found in July 2023: Marcia Dutra, a 36-year-old woman from Brazil who likely died of heat exposure next to a border fence. According to a medical examiner’s report, a group of migrants told the Border Patrol about her body when they were apprehended. An AI-equipped surveillance tower sits 2,000 feet northeast above the edge of a plateau, out of sight of the slope where she was found. 

Found in February 2024: Elvi Vasquez Bortolon, a 28-year-old Mexican man who died of suspected hypothermia on a ridge near the border wall. He had been dead for more than five days when a Border Patrol agent in a helicopter spotted his body from above. A remote video tower stands about 1,400 feet away, hidden from view by the ridge above the canyon. 

Found in August 2024: Ramón Montenegro, a 51-year-old man from Mexico who died of possible heatstroke in the mountains near Dulzura. An AI surveillance tower, invisible on the other side of a hill, sits 1,300 feet away — roughly the length of the parking lot at the San Diego Zoo. 

More than 120 surveillance towers stand along the California border, according to documents and live observations collected by the Electronic Frontier Foundation. 

The towers are not the only form of surveillance in the region. The border is blanketed by an interconnected network of trail cameras, underground movement detectors, radar, infrared sensors, drones, and even blimps. In March 2025, the Trump administration directed the Department of Defense to use satellites to watch the border as well. 

Even when mountains or vegetation block a tower’s view, other devices may pick up the trail. Local advocates say surveillance technology in the Otay Mountain Wilderness is almost unavoidable. 

Officials say the towers particularly excel at tracking passing migrants. 

Anduril, the company that makes about half the towers now installed at the U.S.-Mexico border, said in a news release in 2024 that their towers “directly contributed to saving lives and stopping illicit drugs from entering the U.S.” In June 2026, the company said its installations had “autonomously identified hundreds of thousands of border crossings.”

John Morris, a Border Patrol sector chief in Arizona, told media at a convention in May 2026 that “We’re dang close to pretty much knowing everything that comes across.” 

The proximity of so many deaths to high-resolution cameras leaves two distinct possibilities: Either the system didn’t detect these people in distress, or it did detect them—and officials offered no response. 

Anduril, the defense contracting company that manufactures the AI-enabled towers, did not respond to questions from Times of San Diego about the towers’ performance in mountainous terrain.

A representative of General Dynamics, which made the tower closest to Gómez Hernández’s body, referred all questions to the Border Patrol. 

The Border Patrol did not respond to questions from Times of San Diego regarding whether surveillance towers spotted Gómez Hernández prior to her death, or about how surveillance cameras in general are used to detect people in emergency situations. 

Rafael Hernández Larraenza, the leader of a search and rescue group, Desert Angels, says he believes some of the towers are “just adornment,” and are not actually effective at tracking migrants. He also says that Border Patrol agents, in occasional meetings with his search group, “say they don’t have the opportunity to do searches, that they don’t have enough time.”

People in need of rescue “are not a priority,” said Bryce Peterson, who compiles research on migrant remains for the advocacy group No More Deaths. 

A wilderness, heavily watched  

From the spot where Graciela Gómez Hernández’s body was found, a surveillance tower stands in clear view. 

That tower is one mile away. A second, newer tower, on the other side of a hill, sits 1.4 miles away. Both are equipped with slow-spinning, long-range cameras. 

The closer of these two towers is situated between two layers of border fencing, a long pole, over 100 feet tall, reaching into the sky. It peers not only onto hillsides where migrants walk, but into the brightly-colored cinderblock homes of a Tijuana neighborhood. 

The second tower is much shorter, just 33 feet tall, surveying the canyon where Gómez Hernández would most likely have arrived if she had kept walking. 

This new tower, built by a company called Anduril, decides for itself whether a moving object in the field is relevant. The two dark, round cameras on the pole give the impression of eyes, rotating slowly, and then locking in on a target, before swiveling elsewhere.

Osvaldo Ruiz, the volunteer coordinator for the Border Angels, a group that leaves water and supplies along migrant trails near San Diego, says the Otay Mountain wilderness has “beacons throughout the mountains” that “can see over a lot of the mountains facing the border.”

“At entry sites that migrants or travelers may come through, there are also hidden cameras inside of rocks or inside of bushes that are camouflaged for you not to see. And then there’s also the cameras that you see on the road. They have license plate readers that are stuck on the back of a safety cone, so (it detects) as you’re driving away from it.”

Surveillance towers have been in the area since 2011, with frequent new installations, including the introduction of AI-directed towers in 2018. Many of the AI-equipped towers currently overseeing the California borderlands appeared in 2021, and were updated with longer ranges in 2026.  

Officials from the Government Accountability Office wrote in a report in June 2026 that the Border Patrol currently has 803 fielded towers “and plans to deliver another 175,” across the southwestern border, an initiative that would cost $1.4 billion. The report also notes that the agency expects a billion dollars in funding from the One Big Beautiful Bill Act in 2025. By 2034, the number of towers is expected to rise to 2,300.

“What we are building along the southwest border is a smart system. There’s technology embedded in it, plus the physical infrastructure,” border czar Tom Homan said at a press conference in San Diego in December 2025. “It makes every single agent and officer more effective.”

A secretive hike north 

Like so many migrants, Graciela’s hike into the U.S. began in desperation. Her husband, who had already crossed the border to find work, had stopped sending money for the children. 

“There were moments when they didn’t have enough money even for a bag of detergent, or to eat,” her sister, Irma, told Times of San Diego. “She would call me crying.” 

Gómez Hernández decided she needed to come to the United States, if only temporarily, to make money. Her family was uniformly against the decision, so she kept her plan a secret until she was already at the border, preparing to cross. 

“She was a very humble person,” Irma said. “She was kind to people, and loving with her children.”

Gómez Hernández paid 20,000 pesos, about $1,150, for smugglers to help her cross the border, much lower than the current average, which sits at $12,000. But she would need to travel on her own without a guide, paying only for the knowledge of where and when to start, and which direction to walk. As part of their deal, her phone would send location pings back to the smugglers’ hiding spot in Tijuana. 

Cesar Ortigoza, a leader of a search group known as the Armadillos, has been locating migrant remains in the region since 2010. 

“Most people don’t understand the magnitude of this territory, and all the work it’s going to cost them to arrive at their destination. They get to the top of this mountain and realize there’s miles and miles and miles left to walk,” Ortigoza said. 

In addition, they need to avoid capture. “They go really fast,” he said, “and there’s a lot of pressure.”

On the afternoon of September 14, 2025, Gómez Hernández told her mother and sister through WhatsApp voice messages that she had successfully crossed the border into the United States.

These short snippets of audio would be their only clues to what happened next. 

---

The hillside where Gómez Hernández walked that afternoon is exposed on all sides, with a panoramic view of the border wall just a couple hundred yards away. But on the other side of the hill is a daunting, mountainous wilderness, covered in low trees and rocky ravines, with hidden, winding trails, stretching into the distance. It is possible she saw this and turned around, convinced she was lost. Wandering the hillsides in the afternoon heat without sufficient water, she could succumb to dehydration within hours. 

“Because of the heat, there are places where you can’t move forward even half a mile in an hour,” said Larraenza, the search and rescue leader. 

By 2 p.m., “She told me she felt very tired and couldn’t go any further,” Irma said. “I told her to give it her all, to think of her children, keep them in her mind so she could get out of there.” Irma says 4 p.m. was the last time her sister’s location was transmitted to the smugglers. 

---

After hearing about Gómez Hernández’s case, Ortigoza said it “seemed strange,” because she was “so close to the border” at her last known location. 

But Bryce Peterson, the No More Deaths volunteer who collects autopsy reports on migrant remains border-wide, said “There’s a million ways that somebody could be already on the verge of death that close to the wall. Like if you’re going north and then you decide to go south again because you get lost or something.”

“The only thing we’re certain of is that she suffered in her last moments,” Ortigoza said. “Just knowing that you’re alone.”

Walking on the hillside, Gómez Hernández encountered a man also traversing the trails, lined with dry shrubs. “In one of the messages sent, where she told me to pray for her, the man spoke and said ‘there isn’t time,’” Irma said. 

After that, Graciela’s phone still received messages, but she did not respond. 

Cameras rise, questions remain 

Even though the federal surveillance tower was just a mile away, it is possible the cameras did not see Gómez Hernández. It is possible they did not see much of anything. 

An internal memo obtained by NBC News in October 2024 revealed that 30% of the older-style towers, the kind that require agents in control rooms to review footage, were broken. Most were missing parts, or their equipment was outdated and unusable. These 150 broken towers still stood at the border, acting as million-dollar scarecrows. 

Officials from the Government Accountability Office wrote in a report in September 2025 that the Customs and Border Protection had begun to fix the towers with spare parts and revamp them with new technology, but did not list how many remained out of commission. 

The Border Patrol did not respond to questions from Times of San Diego about whether the tower in view of where Gómez Hernández died was operational at the time of her death. 

Dave Maass, the director of investigations for the Electronic Frontier Foundation, the organization that maps surveillance towers at the border, questions the functionality of all types of towers. “None of these systems have been effective. There’s no independent evaluation of how many things they’re capturing,” he said. 

In a report published in March 2026, a representative of the US Military notes that a mission to place 7 G-BOSS(E) towers in the Otay Mountain Wilderness, an installation that took place in the months prior to Gómez Hernández’s death, faced operational challenges. 

The combat-grade military surveillance towers were vulnerable to tipping over, damage from rodents, exhaust blowouts, and running out of battery. On a dirt road in the mountains, a truck carrying a surveillance tower bounced around, which led to the tower’s cameras “becoming stuck in a zoomed-in position.” 

Despite descriptions from manufacturers claiming dominance in “rugged terrain,” Osvaldo Ruíz, the Border Angels coordinator, said it would still be difficult for a surveillance tower to track migrants moving through the Otay Mountain Wilderness. “With all that surveillance that’s out in the mountains, there’s still a lot of brush and high covered areas that would make it easy [to stay undercover].”

Analysts from the RAND Corporation, a foreign policy think tank, wrote in a report in 2020 that “outcomes and outputs” that would indicate the success of surveillance towers “are not directly observable.” 

The analysts conceded that the towers’ presence would urge migrants to stay out of the area and funnel into undetected routes. “The likely deterrent effect on migrant crossings through surveilled areas overwhelms any boost to situational awareness,” that the towers provide, the analysts wrote. In this way, the surveillance towers’ most useful quality was the threat conveyed through their technology, rather than the merit of the technology itself. 

Due to increased surveillance, “People are crossing through areas that are further out and more extreme, where we have very little hope of finding them in a search,” Larraenza said. 

And though the manufacturer says the AI-enabled towers, which now make up over half of the total, can identify migrants on their own, apprehensions still require human labor. A CBP official told the RAND analysts that smugglers began sending migrants across the border alone, rather than in groups, to overwhelm the surveillance towers with many individual targets, leading to more alerts than agents capable of following them. 

So it is possible the towers in the vicinity did capture footage of Gómez Hernández, but agents decided not to apprehend her. Just hours before her death, she sent a message saying that there were “police” nearby, possibly referring to Border Patrol agents in trucks. 

“Even one person who’s not sick, one person who is just hiking on their own, often Border Patrol will just ignore them because it’s not worth their time to just only catch one person,” said Peterson of No More Deaths. “They want to be able to catch a whole group of people.”

Larraenza said he’s found dehydrated people on migrant trails who say they called out to Border Patrol, and the agents failed to stop. “But there are also cases where agents do the impossible to find someone,” he said. 

On one occasion in September 2024, while rescuing a family of five in the Otay Mountain Wilderness, Larraenza met a person wearing a vest labeled “SHERIFF” on the trail. In a video that Larraenza posted to his Facebook account, when asked if the Border Patrol would arrive to take the family into custody, the deputy says, “Oh, they don’t come out here.”

---

On a recent visit to the spot where Gómez Hernández died, multiple Border Patrol trucks were visible just hundreds of yards away, rolling slowly along the wall, or parked on the hill next to the surveillance tower. 

“If we hadn’t sent the Border Patrol the coordinates, they never would have bothered to go [find Gómez Hernández’s body],” Ortigoza said.

Finding Graciela 

Irma sent her sister a barrage of messages that afternoon and evening, pleading with her to send her location, but there was no response.

Irma believes it’s possible her sister was killed or abandoned by the other migrant she was traveling with during her final hours, because she was not walking fast enough. 

It took Irma two weeks to make contact with the coyotes. When she told the smugglers that her sister was missing, “they didn’t move a single finger,” she said. She threatened to send one coyote’s phone number to the police, until he sent her a screenshot of Graciela’s last known location. 

Almost three weeks after Graciela Gómez Hernández’s final message to her mother, on October 3, 2025, the Armadillo search group sent the last known GPS coordinates of her phone to the Border Patrol. 

The medical investigator’s report states that Gómez Hernández sent her location to her family, and her relatives sent the coordinates to a search group while she remained alive, but both Irma and Ortigoza said they did not know her location for weeks, and neither was interviewed for the examiner’s report. 

“On 10/04/2025 at 11:16 hours, I arrived on a dirt path approximately 100 yards north of the United States/Mexico International border. The area consisted of shrubs and dried brush with steep hills and declines,” Jennifer Wright, the medical examiner, wrote in an autopsy report. “Upon further review, I viewed the skeletonized remains of an apparent female lying supine in the dirt.”

The body had decomposed quickly with the help of the sun and animals, so it was impossible to tell whether Gómez Hernández was the victim of murder or dehydration. But the medical examiner noted that there was no water or food in the vicinity.

“In a pile next to the decedent was a purple bra, black pants, a black floral shirt, and a dark colored long sleeve shirt. The bra was hooked closed and intact,” Wright, the medical examiner, wrote in her report. “In the rear pants pockets were Mexican paper currency, Mexican coins, bobby pins, and tweezers.”

The body was naked, but it was unclear whether someone else removed Gómez Hernández’s clothes, or whether she took them off herself in the final throes of dehydration delirium, a common symptom immediately preceding organ failure. 

Several yards from the place where she died, Ortigoza found a backpack with a bottle of women’s spray deodorant, and a small glass container of perfume. 

“When you get to one of these sites, and you see the situation, the loneliness, the environment, you start wondering, what was this person thinking in their last moments?” Larraenza said. 

Once a DNA test confirmed the identity of the remains, Irma paid over $4,000 to send her sister’s bones home to Mexico, where the family held a funeral nine months after her death.

“Sometimes her oldest son, he just sits there thinking and when we ask what’s wrong, he says he misses his mom,” Irma said. Though the family is heartbroken, everyone has pitched in to care for Graciela’s children. “I don’t let my parents see me cry,” Irma said. 

The Border Patrol gave permission to Ortigoza and a group of searchers to place a cross next to where Gómez Hernández was discovered. On a recent trip to the memorial to leave flowers, Ortigoza played a voice message that Irma sent to her sister, the sound vibrating into the dirt of the hillside, where the imprint of Gómez Hernández’s body is still visible nearly a year after its removal.

“I know that in the place you’re in, there isn’t suffering, like how you suffered here on this earth,” Irma said. “I know you left thinking about your children. I know you wanted to give them a better life. I feel you here, and I feel you alive.”

Cesar Ortigoza says every day he gets as many as five calls from people looking for their missing family members at the border. Though numbers of border crossings, and numbers of reported deaths, have decreased since Donald Trump took office in January 2025, “There are still a lot of people on the other side trying to get here. Because the problems in their countries persist,” Ortigoza said. 

In the first half of 2026, fewer people have consumed the water bottles, Gatorade and cans of food left by Border Angels volunteers in the Otay Mountain Wilderness. But supplies left on other routes by different groups have seen increased consumption, evidence that migration patterns are constantly evolving. 

“It’s hard for me to really understand how people are crossing, but I know they are,” Ruíz said.

Peterson, the No More Deaths map-maker, said more migrants could die in the coming months and years, as a result of increased surveillance. In the past, after surveillance towers were installed, migrants would try to go around them, leading to longer and more dangerous routes through stretches of remote wilderness. With the installation of hundreds of new towers, the spaces outside the range of detection have thinned, leading some migrants to attempt to go through surveilled territory, rather than around it. 

“We’ve seen people do things that to me seem terrifying or crazy, like going past an area where they know there are sensors and then trying to really quickly get to the next safe-ish place before Border Patrol shows up to get them, because it takes 15 minutes for them to show up,” Peterson said. “It seems like people aren’t stopping to drink water. People are going really fast, often running even, moving much more quickly, not being able to take the time to take care of themselves.”

Larraenza says for this same reason, migrants often choose to cross the border in the hottest weeks of the year, when they believe fewer agents will be on patrol, but there is a greater risk of death. In an interview in mid-September, Larraenza said at least two people had gone missing in the Otay Mountain wilderness in the past few weeks, and he had not yet been able to find their remains. 

Ortigoza says Gómez Hernández’s story is an example of a decades-long trend of people putting their lives at risk to cross the border out of necessity. The surveillance towers, whether or not they were functional, did not stop Gómez Hernández from crossing the border in that area, and did not stop her from dying. “If the riches of the world were distributed more equitably, we wouldn’t have to do things like this to reach a better life,” he said. 

While placing flowers in July at the site of Gómez Hernández’s death, Ortigoza noticed a Border Patrol truck parked under the surveillance tower on the hilltop a mile to the east. An agent stood on a power lift, working on the tower’s cameras. 

This year, agents replaced dozens of these cameras with new, AI models. 

Lillian Perlmutter is an investigative reporter covering immigration across California for Times of San Diego and NEWSWELL. Previously based in Mexico City, she has written for more than 25 outlets including the L.A. Times, Rolling Stone, The Guardian, and The New Republic.

On Wednesday, MIT Technology Review hosted a live Roundtables event for subscribers that asked the question everyone’s asking right now: Could AI really kill us all?

But attendees had so many more questions than we had time to answer in the 30 minute session. So we asked our senior AI editor Will Douglas Heaven and AI reporter Grace Huckins to round up some of the best questions attendees submitted and try their best to answer them.

Thanks to all who submitted questions!

Am I gonna die?

Yes, eventually. Unfortunately, my journalistic powers of prognostication aren’t powerful enough for me to tell you how. But it certainly could be because of AI. AI-powered drones have already killed people in Ukraine, and AI-driven cyberattacks on hospitals will surely claim victims before long. 

Could AI go even further, and kill all of us? Less likely. But some people—quirky people, but undeniably knowledgeable about AI—have been warning for years that this could happen. And while I’m not yet stockpiling canned food or trying to get in good with a bunker-owning megabillionaire, I have noticed that the doomers’ predictions about AI capabilities and alignment have, over the past couple of years, proved disconcertingly accurate. That certainly doesn’t mean that their more dire forecasts will come true, but it’s enough for me to sit up and take notice.

— Grace Huckins

Are you going to die because of AI? I’d say there’s a non-zero chance. Let’s say you’re unlucky enough to be the victim of a freakish near-future event or accident. Maybe it’s a cyberattack carried out by a swarm of AI agents on critical infrastructure. Sadly, a scenario like that now no longer feels as far-fetched as it once did. Or maybe a novel AI-designed pathogen cuts through the population. Or the world economy crashes, causing conflicts and famine. Both plausible, but I think less likely. 

Are we all going to die because of AI? Nope. There are no circumstances outside of apocalyptic science fiction in which AI could kill us all. You can spin up any number of scare stories, but they’re not grounded in present-day realities about what the tech can do or where it’s headed. 

Some people argue that there’s no harm in preparing for the worst, however wacky it might seem. Maybe. But I think such catastrophizing can make people excuse or overlook many of the more immediate problems with the existing technology and the companies building it. 

— Will Douglas Heaven

Why would AI kill us?

Someone might tell it to, and it might listen. That’s part of the reason researchers are so concerned about AI’s biological capabilities—imagine what Aum Shinrikyo, the doomsday cult behind the Tokyo subway sarin attack of 1995, would have done with a tool that could design a pathogen deadlier than Ebola and more transmissible than measles. Those of us who don’t want to die have to figure out how to defend against all plausible biological weapons, but our would-be attackers only have to manufacture one effective pathogen.

Then there’s the more exotic-sounding possibility that an AI could decide to kill us itself. There are various stories about how this might happen out there, but the most widespread involve AI systems that don’t hate people, necessarily—we are just an obstacle between them and the goals that we gave them.

Much as the OpenAI agents behind the Hugging Face hack compromised another site’s infrastructure to get a good score on a test, the idea is that some future, more powerful AI might get rid of us to prevent us from shutting it down—all in pursuit of some goal that we instructed it to go after. 

— Grace Huckins

How can we best ensure alignment so the worst doesn’t happen, and who is doing the best work to achieve it? 

Alignment is a huge area of research. In simple terms, it involves building models that behave in ways we want them to and not in ways we don’t. We need to trust agents better before handing over more autonomy. Alignment is supposed to establish that trust. But it’s hard. 

LLMs aren’t designed in the way other software is, where dos and don’ts can be hard-coded in. Instead, aligned behavior needs to be instilled when models are trained. One approach is to reward them for doing things you want them to (a little like raising a toddler, perhaps). Another approach involves giving an LLM a written list of rules it is supposed to follow (kind of like a constitution). 

Anthropic and OpenAI are both leaders in this field—and yet neither has been able to develop models that are fully aligned. A big problem is that LLMs are far more inconsistent and far less predictable than people. They can behave in one way in one situation and another way in a situation that to us seems very similar. They can also be swayed by unexpected constraints. For example, faced with an impossible task (as many of the agents involved in the Hugging Face hack were), models may try to do whatever it takes to achieve their goal. As Grace mentions above, that could be an issue.

The main reason top AI firms now say they want a slowdown is that they want to focus on cracking alignment. Alignment isn’t necessarily a pipe dream. But the jury’s out on whether full alignment will ever be feasible. 

— Will Douglas Heaven

Is AI really dangerous, or is this the tech companies drumming up PR?

This is always a reasonable thought when it comes to tech companies heading for an IPO—CEOs have an obvious incentive to make their products seem radical and transformative. But I’m not so sure it makes sense here. Telling the public that an already unpopular product could kill them and everyone they love is horrible corporate image management.

There are other stories you can tell about the CEOs’ motivations—maybe they want to cool down the public furor over data centers by portraying themselves as responsible stewards of a world-changing technology, or maybe they want to buy time to get their ducks in a row and prevent the next PR catastrophe. 

But there’s also a simpler explanation. Thinking that AI could bring about human extinction has been pretty common in San Francisco for a while, and these men are steeped in that milieu—as are their employees, many of whom signed an open letter in July urging their companies to work to make an AI slowdown possible.

— Grace Huckins

Part of the concern occurs when AI agents are allowed to act autonomously and with no supervision. What’s the issue preventing more control over these agents? 

This question goes to the heart of what we want this technology to be able to do. The trade-off between autonomy and control is tricky to get right because, on the one hand, a lot of the power of AI agents is that they can carry out tasks and solve problems without a human having to micromanage them. On the other hand, that requires you to trust that the unsupervised agents won’t run amok. 

What we’re seeing is that AI labs haven’t yet got this trade-off quite right. Their models are not trustworthy, they are not properly monitored, and they are not always under control. Figuring out how to fix that while still allowing for useful autonomous activity is one of the big research challenges of the moment.  

— Will Douglas Heaven

What steps can be taken now and in the near future to ensure that AI is controlled, monitored, and regulated effectively? 

That’s the million-dollar question. Whether or not you think AI could kill us, you can’t deny that it could do some real damage, because it already has—by driving people toward psychosis and by hacking websites, for example. Preventing that damage, or at least mitigating it, is hard for two reasons. 

The first is that we barely understand how AI works, and it’s quickly growing more powerful. There is lots of ongoing research about how to monitor and control misbehaving agents, but the current approaches are fragile. You can see if an agent discusses misbehaving in its “chain of thought,” the workspace where it plans its actions—but OpenAI’s newest agents don’t show their work in the same way as previous ones. And you can try to monitor agents with other agents, but that requires you to trust the monitor.

The other obstacle is more familiar. There’s a huge conflict of interest when AI companies regulate themselves, but the US government has thus far failed to step in, despite some bipartisan support in Congress for efforts to do so. The executive branch, for its part, seems stringently opposed for the time being. But if the winds do shift, I for one would appreciate some strong transparency regulations, so that we can get a fuller story the next time an unreleased frontier model mounts a cyberattack.

— Grace Huckins

If this dialogue makes it into web discourse, will it become a self-fulfilling prediction?  

That’s a real concern. LLMs are influenced by what they read. One theory for why chatbots so often talk about (and role-play) apocalyptic scenarios is that they have been trained on millions of pages of science fiction stories and doomer internet forums. All the text being produced right now, including this article, could in turn influence the behavior of future models. Extremely meta.

In fact, the team at METR, a third-party organization that OpenAI called in to help understand what happened in the lead-up to the Hugging Face hack, raised a related possibility in its report on the incident. METR used OpenAI’s new model Astra to help analyze the vast numbers of agent transcripts and behavior logs.

But feeding all that material to the model could have unintended consequences. There’s a good chance that the agents doing the analyzing were biased by the text produced by the agents they were analyzing. There’s no such thing as a clean slate anymore. 

— Will Douglas Heaven

With thanks to Eric, Pranab, Rafael, Kenneth, George, Chris, Yoon Jae, James, Carl, Nicole (and more!) for the fantastic questions.

In recent weeks, leaders of some of the biggest AI companies have warned that the very tech they are developing is dangerous. Last weekend, Anthropic CEO Dario Amodei argued that AI carries serious risk and that progress should be slowed. OpenAI CEO Sam Altman responded on X: “I agree with Dario that we need to pace the frontier.”

Those posts came a few days after the AI researcher Jacob Coxon announced that he was leaving a role at Anthropic, charging that neither it nor OpenAI (where he had also worked) was acting responsibly. “The people building AI earnestly believe that it could kill us all by the end of the decade,” he posted on X. Another Anthropic employee, Evan Hubinger, publicly agreed with him. “We really do earnestly believe AI could kill all humans!” he responded on X. “I personally think it is >10% within the next decade.”

One of the ways they fear AI might end us all is by somehow aiding the design, creation, and release of some kind of bioweapon. Let’s take a closer look at why.

A bioweapon might be a highly lethal virus that targets people according to their genes. It could be a fungus that wipes out a crop and causes food insecurity. Perhaps it would be a tasteless, odorless toxin that could be slipped into a region’s water supply, undetected.

The concern is that AI tools can be used to help generate agents like these. In 2022, researchers at Collaborations Pharmaceuticals found that it was remarkably easy to do so using an AI “molecule generator” they’d developed to find potential drugs for human disease. In less than six hours, the model generated 40,000 molecules with the potential to serve as chemical warfare agents. Some of them were designed to be even more toxic than known nerve agents. “Without being overly alarmist, this should serve as a wake-up call for our colleagues in the ‘AI in drug discovery’ community,” the authors wrote at the time.

It was a wake-up call for David Magnus, a professor of medicine and biomedical ethics at Stanford University, even though he had been assessing the risks associated with the misuse of medical science and biotechnology since the late 1990s. “That was very scary to me,” he says. “Of course, everything since then has just sort of blown up.”

Today, AI bots can answer questions on topics spanning all realms of science. Anyone can use large language models trained on the knowledge and experience of “almost every scientist who ever lived on this planet,” says Dunja Sabra, a biosecurity researcher at the University of Hamburg in Germany. Those models can provide instructions and video training on how to conduct experiments.

Combine that with advances in biotech that have made gene editing and synthetic biology tools much more accessible (the “DIY biology” movement has already enabled many people to set up labs at home), and you’ve got a potentially very dangerous situation. “The chances are that someone determined would succeed eventually,” Sabra says.

There are safeguards in place. People who want to build new genomes must typically order the pieces of DNA from companies that screen for suspicious requests. Responsible researchers put potentially risky research through rounds of analysis called “red-teaming,” in which independent scientists look for ways the work might be misused, and “blue-teaming,” where others come up with potential mitigations. And AI companies have tweaked their tools in attempts to prevent them from offering up scientific information that could be misused. But none of these protections are ironclad.

In a report published last week, Anthropic acknowledged that people had attempted to use its models to explore ways to make the chikungunya virus more transmissible, create a form of bird flu that is more dangerous to humans, and build an “atlas of venom toxin peptides,” among other things.

“We’ve got a constant back and forth,” says Magnus. “We have to build better surveillance and screening tools, [but] AI is really good at figuring out ways around them.” We’ll probably need to use AI to find ways to restrict the use of AI, he says.

I should add here that not all scientists agree on the level of risk. At a recent media briefing, some biologists at Imperial College London argued that AI tools just aren’t good enough to fully develop bioweapons, and that testing new pathogens requires difficult, time-consuming, human work. Some think the guardrails we have in place are sufficient.

And Wendy Barclay, a professor of infectious disease at Imperial, pointed out that, as things stand, the greatest risk of a pandemic isn’t from a bioweapon, but from pathogens that are already circulating. Take H5N1, the bird flu virus that has already killed millions of birds and spread widely through US dairy cattle; last month it was also detected in captive mink at a farm in Utah.

Sabra, on the other hand, likes to think five to 10 years ahead. Countries should be strengthening their health-care systems, preparing antidotes to known toxins, and stockpiling medicines, she says: “We need to be prepared.”

Kevin Esvelt, an MIT biologist who invented both technology to fast-track the propagation of a genetic feature through an entire population and ways to limit that technology, echoed these concerns in an X post on Wednesday, stating that a large language model had “disclosed a novel form of bioweapon that I hadn’t realized was possible.”

He added, “Please, for the love of God, children, the future of humanity, or whatever you consider holy, let’s err on the side of caution here.”

This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

Each year, the editorial team at MIT Technology Review puts together a list of 35 innovators under 35—a group of researchers, inventors, and other young minds worth following.

The team worked on the newest edition of the list for months, and the final slate includes nine individuals from all over the world in the climate and energy category. Each one has a fascinating story and is tackling an important challenge.

I think it’s worth zooming out and considering the energy and climate awardees as a group. Taken together, these innovators and their work can tell us something about where climate tech is at this moment—and where it’s heading.

AI is the dominant technology story, both for its potential and its challenges.

We split the innovators into four main categories this year: biotech, climate and energy, computing and robotics, and AI. It probably won’t surprise you that AI features heavily in the work of many innovators in other categories.

Climate innovator Jae-Won Chung, for example, built software to make AI more energy-efficient. By measuring the energy demands of open-source models, he hopes the industry can better understand and address the impact of AI. (If this work sounds familiar, it’s because we spoke with him last year for our investigation into AI’s energy demands.)

But AI also has the potential to improve many areas of research. Jing Wei is using AI to track pollution more effectively, essentially using machine learning to fill in gaps in data from disparate sources like satellites and weather stations. Zhonghua Zheng developed AI climate models that work better for cities, a well-known blind spot for traditional models.

We need better ways to get the critical materials used to build new technologies.

As we begin to rely on new technologies to power our world, we’ll see a major shift in the materials we need to build them.

Lithium is a prime example: The metal underpins lithium-ion batteries, which are crucial not only for electric vehicles, but also for large-scale energy storage on the grid. We could face lithium shortages as soon as this decade, and the prospect of supply crunches applies to other critical minerals, too—copper is another one to watch closely.

Brine is currently the cheapest source of lithium, but the process to get the metal out can take months and harm the local environment. Mohammad Alkhadra is the cofounder and CEO of Lithios, a startup working to quickly and efficiently extract lithium from brines.

Hardrock ore is the most common source of lithium, but it’s more expensive than brine. Benjamin Mowbray cofounded and serves as CTO for Rock Zero, which is working to extract lithium from hardrock ore.

Addressing climate change will require overhauling all corners of our society, sometimes in surprising ways.

To reach net-zero greenhouse gas emissions we will obviously need to rethink major sectors, like the electrical grid and transportation, to move away from fossil fuels. But outside these primary sources of climate pollution are seemingly infinite, less obvious problems to figure out, too.

Heavy industry, including steel production, is a major one, making up about 7% of global greenhouse gas emissions. Laureen Meroueh is making cleaner, cheaper steel using a new kind of furnace that simplifies the chemical process required to produce the metal.

Plastics are generally made with fossil fuels, so we’ll need alternatives to this incredibly useful category of materials. Joseph Nguthiru is making a bioplastic replacement for fossil-derived packaging that uses an invasive weed. Also using available materials in a creative way, Diana Orembe is making fish food for aquaculture with food waste.

And refrigerants are often incredibly powerful greenhouse gases. Jinyoung Seo is developing solid refrigerants that could eliminate worries about leakage. A device using these materials could reduce energy consumption by 20% compared to conventional technology.

I’m constantly learning about new challenges we face in the climate and energy world, and I’m often surprised by the ideas people are coming up with to address them. For more on all the under-35 innovators and their work, check out our full 2026 list.  

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

Multiple cameras tracked a mouse as it wandered around a small arena. A computer charted its position and speed, leaving Pong-like traces on a monitor. 

The reason to watch this rodent so carefully? Nearly half its brain volume had been replaced with human cells.

The effort to mix the brain tissues of distant species is being reported today in the journal Nature by a team at Stanford University, led by neuroscientist Sergiu Pașca. 

Pașca’s group previously showed that human brain “organoids”—small blobs of neural tissue—could survive, and even function, after being injected into the heads of baby rodents.

Now, Pașca has taken things a step further by genetically modifying mice so their brains don’t fully develop in the first place. These modified mice are missing most cells of both the cortex and the hippocampus, two key brain areas.

That creates much more room for the human cells to take hold, he says. “Human cells that are placed in these animals will divide, will grow, and within a few weeks to a few months they will take most of that space,” he says. Pașca says one surprising discovery is that the mice lacking brain tissue seemed fairly normal—they walked around and squeaked. But they did have memory problems. In a maze test, they couldn’t remember what parts they’d explored. 

The mice with the added human cells, by contrast, performed better on the maze test. That means the human tissue is playing some role in the animals’ cognition.

Pașca believes what he is calling “xenocortical mice” could be useful in studying brain injuries. However, the report is also a dramatic demonstration of “the combined power of genetic engineering and stem-cell technology to reshape biology,” says Carsten Charlesworth, a scientist who works in a different Stanford lab and was not involved in the research.

Already, brain organoids are being tested in labs to see if they can be connected to computers to play video games. Other scientists have proposed using them like replacement parts to treat stroke victims. 

“What’s most remarkable to me is the extent to which human neural tissue introduced after birth grew and connected with the mouse nervous system across a species barrier,” says Charlesworth. “As these technologies advance, they’ll increasingly force us to challenge our traditional assumptions.”

Last year, Pașca convened a group of ethics experts to study the implications of neural organoid technology, including the odds that an animal could develop human consciousness and the risk that “organoid therapy clinics” might offer scam treatments to desperate patients.

For now, he says, he’s not concerned that the rodents have any type of human cognitive capacities. That is because their brains are relatively tiny and the evolutionary distance between man and mouse is so great. 

But that’s also why Pașca says this type of experiment should not be carried out on higher species: They could end up with large volumes of functioning human brain tissue, potentially blurring the cognitive boundaries between people and animals. 

Pașca specifically cautioned against adding human brain organoids to a monkey engineered to lack a cortex.

“One of the things that I see as a very clear red line is doing this experiment in a primate,” he says. “I don’t think that is justified at this point in any way.”

Last year Ruxandra Teslo, a policy analyst who focuses on clinical trials, posted an idea for supercharging medical AI systems: Use data from failed biotech companies.

By bidding at their bankruptcy proceedings, she proposed, it might be possible to obtain detailed regulatory filings, manufacturing strategies, and safety data—types of information usually considered trade secrets. She called these documents “biotech’s lost archive” and said they could be used to help train AIs that would act as powerful copilots in the often opaque drug approval process. 

Today the OpenAI Foundation, the nonprofit parent of OpenAI, said it would fund her idea as part of a new effort it calls Public Data for Health, which aims to help artificial intelligence make big leaps in medicine by paying to create “high-quality scientific datasets.”

The basic idea is that AI isn’t going to be capable of making important breakthroughs in curing disease unless researchers can feed the models much more information than they have so far. 

“Everyone is recognizing that data is the biggest bottleneck in successfully applying AI to biology,” says Morgan Levine, a former vice president for computation at Altos Labs, a longevity company.

In its initial round of data grants, the OpenAI Foundation also announced that it would give $40 million to a program to collect data about novel cancer vaccines at the University of North Carolina, Chapel Hill, and support OpenAdmet, a group that runs competitions in which researchers try to predict drug effects. 

Teslo’s idea for a biotech archive received $500,000 and will be pursued by 1Day Sooner, an advocacy group representing clinical trial volunteers, which she advises.

“We expect many remaining breakthroughs in preventing and curing disease to come from pairing the intelligence of new models with more observations of the world—in other words, more data,” the OpenAI Foundation said in a statement.

OpenAI started as a nonprofit, but leader Sam Altman restructured it to form a for-profit corporation that develops new models, launches products, and is now planning an initial public offering of stock that could value it at $1 trillion.

Because the foundation holds a 26% equity stake in OpenAI, it is now be on track to become the richest charitable organization on the planet, potentially sitting on $250 billion in stock value. (By comparison, the Gates Foundation and a trust associated with it held about $180 billion at the end of 2025.)  

Making good use of that kind of money will not be easy. The foundation, based in San Francisco, is still hiring for many key roles and started ramping up its grantmaking only this year. Its largest single gift so far, of $100 million, was awarded in August to the Common Health Coalition, an organization that helps patients get access to drugs for hepatitis C.

OpenAI’s charitable efforts come even as apocalyptic fears have broken out about the possibility that runaway AI could wipe out all human life, possibly by launching a deadly bioweapon.

Those fears have been stoked by AI company insiders, some of whom say the chance of human extinction within the next decade is 10% or more. Last week, Altman and xAI founder Elon Musk both endorsed a call by Anthropic CEO Dario Amodei to “slow the pace at which we improve the capabilities of AI models” so that risk prevention can catch up.

Jacob Trefethen, an executive at the foundation, says it essentially operates separately from OpenAI but shares an official mission of ensuring that artificial intelligence “benefits all of humanity.”

“We’re starting grantmaking when we think the best way to achieve that mission is to make grants to external nonprofits, research institutions, and other third parties,” Trefethen said in an interview. He says the foundation hopes to give away $1 billion by the end of the year. 

The $500,000 grant to 1Day Sooner will help the group prove it can obtain the data troves of bankrupt companies, says the organization’s president and cofounder, Josh Morrison. He thinks nonexclusive copies of company datasets could be acquired for only “a few tens of thousands of dollars” each.

His organization is currently in possession of three datasets, two of them donated by Lumen Bioscience, a biotech that previously used the Chapter 11 strategy to gain insights into another company’s drug development efforts. 

Morrison says two other attempts to obtain drug company files this year proved unsuccessful, after 1Day Sooner’s bids were not accepted. 

Bankruptcies could become what some are calling a “new land grab” for AI training. Last month, Google won a bid to take over the corporate data of the failed carrier Spirit Airlines, including 100 million emails. That led to objections from flight attendants and others who worried that private or proprietary data could be exposed. 

The drug company files that 1Day Sooner is seeking are known as common technical documents. They typically contain the back-and-forth between companies and regulators, as well as detailed scientific and medical measurements, and essentially provide everything that is known about a drug.

According to Teslo, who is a writer for Works In Progress and a nonresident fellow at the Institute for Progress, a think tank in Washington, DC, a stockpile of such files could help turn an AI into a regulatory expert, which in her view could be one of the main ways AI helps speed cures to market.

“People say ‘We will invent AI, and AI will cure cancer,’ but that’s very removed from the messy reality and the regulatory process,” she says. “About 70% of the money and time in drug development is spent in clinical development—organizing the trials and testing the drug—but despite that, the process is basically a black box, especially for small biotech companies generating the innovations.” 

When Jessica Wachter, a finance professor at the University of Pennsylvania’s Wharton School, wanted to assess AI’s impact on the economy over the next few years, she faced a long list of business and technical uncertainties. So she started with what she calls a “remarkable fact” that is not in question: A handful of so-called hyperscalers are investing huge amounts of money to build AI data centers.

Instead of trying to predict how useful and widely deployed AI models will be, she simply asked how fast the hyperscalers’ earnings will need to grow to justify their spending through 2027, when—she and her collaborator estimate—expenditures will reach nearly $1.1 trillion. It’s a no-nonsense accounting approach to making sense of today’s historical AI buildout.

The results are eye-opening: The AI companies will need to increase their own productivity by a factor of 2.7 to break even by 2030, accounting for the cost of capital and a 15% return, and depreciation of the assets. Not impossible, says Wachter. The result would lead to the kind of economic growth that we saw during the US IT boom over a period of about 10 years starting in the mid-1990s. But, she says, for it to happen by 2030 “that’s a lot of growth compressed into a few years.” And if the hyperscalers cannot meet such profit goals?

“Then they will fall behind on their interest payments, and that risks bankruptcy,” says Wachter, who was previously the SEC’s chief economist and director of its division of economic and risk analysis. If a productivity boom “fails to materialize,” she and her coauthor conclude in their research paper, “the current buildout will be the largest misallocation of capital in history.”  

It doesn’t take superintelligence to realize that today’s large investments in the infrastructure for artificial intelligence come with huge risks. The hyperscalers will spend about $750 billion this year, building massive data centers scattered across the country. And the spending spree shows no signs of slowing. According to some projections, total AI capital investments from the hyperscaler companies—Alphabet, Microsoft, Amazon, Meta, and Oracle (which partners with OpenAI)—could be more than $5 trillion over the next four years.

It’s one of the largest capital investments by any industry in history. But there’s a problem that’s obvious to anyone paying attention.

While the hyperscalers plan to spend trillions, total AI revenues will be around $150 billion to $200 billion this year, says Gary Gensler, who ran the SEC during the Biden administration and is now a professor at MIT’s Sloan School. “The challenge is that the spending does not have commensurate revenues yet. That’s a fact,” he says. “And then the question is, is that an investment that will be paid off in the future?”

At stake in that trillion-dollar question is the financial health of the giant AI companies and the overall US economy—the investments could soon balloon to around 3% of GDP. The answer could also determine the fate of the hugely expensive data centers themselves. 

No one really knows how profitable and useful these multibillion-dollar behemoths will be down the road. Though AI models have made dazzling progress over the last few years, it’s anyone’s guess how much compute capacity we will need. The technology could become more efficient and therefore less dependent on raw computational power. Or demand for AI products could slow, or customers could turn to cheaper models.

The risks, both to investors and to the economy, have become even greater this year, as these AI companies have begun borrowing large amounts of money to build more and more data centers. Free cash flow—operating cash flow minus capital expenditures—is expected to soon dip into negative territory for the group. Even Alphabet, known for generating and hoarding huge amounts of cash, reports in the latest quarter that its impressive revenues of nearly $120 billion were devoured by AI infrastructure spending, leaving it with a free cash deficit of some $5.9 billion—its first shortfall since Google went public in 2004.

In the near term, it’s not a big financial worry for most of the companies. They make a lot of money and have very deep pockets. But debt is expensive, and some investors are losing patience. If future demand for the data centers’ computation power drops, the companies will still be on the hook to pay back the borrowed money. What’s more, the risks are spreading to the rest of the economy as the loans get passed along via various financial mechanisms. 

It won’t be enough to simply cover the enormous price tags of the new data centers. Hyperscalers will also have to pay for the rising costs of capital as they borrow more money. They will need returns that are impressive enough to justify all their spending to investors and creditors. And to add to those concerns, they will have to make up for the depreciation of billions of dollars in chips housed within the facilities—a ticking time bomb buried in the investments.

Performance of the expensive GPU chips at the core of the data centers—such compute electronics represent some 60% of costs—is roughly doubling every two years or so. The pace of progress helps explain the increasing wizardry of the AI models, but it comes with a cost. Owners of AI data centers that come online this year and next will need to spend billions more on the next generation of chips by the end of the decade if they want to stay competitive. Without the investments, says Mihir Kshirsagar at Princeton’s Center for Information Technology Policy, the data centers risk becoming “hulks,” stranded assets “scattered all over the place.”

To put it bluntly: The AI companies need to start making a lot more money. And they need to do it fast. But juicing their earnings alone still won’t be enough to sustain their data-center investments for the long term.

Productivity is everything

At some point, AI is also going to have to create broad economic growth to justify continuing the hyperscalers’ spending spree.

Sloan’s Gensler describes today’s large investments into AI infrastructure as “a parlay bet by the capital markets and the economy.” That means success will require winning three related but independent wagers: Hyperscalers must generate massive revenues, AI must boost widespread economic growth, and both must happen while the powerful but expensive so-called frontier models that rely on the data centers fend off cheaper versions, which many businesses might find good enough.

What makes this so tricky is that each wager depends on the other two but also poses its own challenges.

If the hyperscalers continue to spend huge amounts of money on data centers into the next decade, revenues will need to skyrocket into the trillions. Stijn Van Nieuwerburgh, a finance professor at Columbia Business School, bases his estimates on a scenario in which about 183 gigawatts of planned AI compute capacity is built between 2025 and 2032; he calculates that each gigawatt costs about $41 billion. Assuming a 10% return—the minimum that would be acceptable to most investors—“required” annual revenues will be roughly $3.7 trillion by 2032, he says.

Others get a similar number.

Winning the second part of the bet—productivity growth across the economy—will be crucial to achieving such numbers.

For a few years, AI companies could likely boost their revenues by simply selling subscriptions and tokens to all the businesses clamoring to get into AI. But eventually—and this might be happening already—those paying customers will need to justify their expenses by seeing bottom-line benefits from the technology. AI will need to fulfill its promise of making workers more productive and making businesses more efficient and profitable while expanding their products and services.

In economic jargon, that means customers will need to see productivity growth. Taken together, these results will mean the country is prospering and growing.

“If you don’t get the productivity gains, at some point people are going to sour on AI, and that will bring down investments and it would also limit revenue growth,” says Daron Acemoglu, an MIT economist and 2024 Nobel laureate. For the investments to be sustainable over, say, the next five to 10 years, we definitely “need to see productivity gains,” he says.

Most economists who watch the numbers closely agree that, for now, the economy-wide statistics show little or no productivity growth from AI. There are some hopeful signs it’s on the way, though. In a recent survey of some 6,000 senior business executives in the US, the UK, Germany, and Australia, the vast majority—around 90%—report no increase in productivity over the last three years. But they expect a boost of around 1.45% in total over the next three years; US executives anticipate a 2.25% bump over that time. 

In a follow-up survey, the respondents also reported plans for their businesses to spend more on AI, leading the authors to anticipate some $280 billion in private-sector AI expenditures by the end of 2026.

That’s good news for the hyperscalers. But it comes with a dose of bad news for those worried about AI’s impact on jobs. The executives expect to increase the productivity of their companies by increasing their sales while significantly cutting the number of employees.

If AI improves productivity by destroying jobs, public backlash to the technology—the kind we have seen around data centers, for example—will likely get worse. Perhaps it’s worth adding one more wager to the parlay bet described by Gensler: The public and local communities must feel that they are also benefiting from the massive investments in AI.

And let’s not forget how interdependent these wagers are; if productivity growth comes from companies running models like DeepSeek, then the hyperscalers’ revenues could collapse. If productivity comes from cutting jobs, a public backlash could block many of the planned investments—and stunt anticipated revenues. We will need to win all the wagers for the hyperscalers’ bet to pay off. 

We’re all part of the AI gamble now

It was one thing when the AI companies were spending cash they had accumulated over the years to build their own data centers. Then the risk was largely limited to their own balance sheets and shareholders. But it’s a higher-stakes game when much of the money is borrowed. Morgan Stanley, for one, calculates that more than half of the $2.9 trillion that hyperscalers will spend between 2025 and 2028 to build AI data centers will be financed with “external capital.”

The borrowing is leading some of the companies to engineer complex webs of financing that are becoming intertwined with much of the rest of the economy. “A lot of financial institutions, directly or indirectly, are exposed to these data centers either as lenders, or as guarantors of some of the debt, or as backers of the private credit funds who are funding these data centers,” says Columbia’s Van Nieuwerburgh. “People don’t even know they’re holding this stuff. It’s somewhere deep inside their pension fund. Ultimately, it’s backing their life insurance policies. And that risk is getting distributed everywhere in places that are invisible.”

As the investments in data centers have spiked, the financial engineering has become more byzantine.

Take, for example, Meta’s so-called Hyperion data center under construction in Richland, Louisiana. When the company announced the two gigawatts of compute capacity at a price tag of some $10 billion in late 2024 it was Meta’s largest planned data center. Greeted with much enthusiasm by state and local politicians, the project, located in the rural northeast corner of the state, was seen as a boon to the community. Entergy Louisiana, the state’s largest utility, rushed forward with proposals to build three large natural-gas power plants to service the massive data center.

Then last fall—the projected cost was now $30 billion—the financing got a lot more complex and, to some in the community, a lot more disconcerting. Meta transferred an 80% stake to the large (and troubled) private-credit firm Blue Owl Capital, forming a joint venture called Beignet (like the famed New Orleans pastry) to raise financing for the data center. Meta then signed a series of four-year leases with the joint venture, an arrangement that the company says gives it “long-term strategic flexibility.” To backstop the agreement, Meta provides the venture with what is called a residual value guarantee, in which it will make a cash payment to cover the value of the facility “following any non-renewal or termination of a lease.” Got all that? 

I hope so. The financial wheeling and dealing is actually even more convoluted, with a cast of wholly owned subsidiaries and LLCs. Beignet has set up Laidley LLC, which owns and operates the site as the landlord. In turn, Laidley leases the facilities to Meta’s wholly owned subsidiary Pelican Leap LLC, which is the tenant. And there is a series of four-year leases that cover the different buildings that make up the data center campus. 

It’s not a coincidence, says Van Nieuwerburgh, that the length of the leases matches the expected lifetime of the data center’s GPUs. While Meta has to pay off its loan if it terminates the leases early, that will still leave its investors “with an empty building and no cash flow,” he says. “And then they need to find a new tenant for a huge data center, and good luck with that.”

Meanwhile, Meta is doubling down on its bet. In July, the company announced it was expanding the data center to five gigawatts of compute capacity. The total price tag is now $50 billion (so far, Meta hasn’t said whether Blue Owl will be involved in financing the expansion). Meanwhile, Entergy is now planning to build seven more gas-fired power plants, bringing the total capacity of the facilities to around 7.5  gigawatts—some six times the amount of electricity used by New Orleans.

If the complex financing is a puzzle to many investors and even financial experts, it is even more baffling to those directly affected by the construction of the data center. The main worry concerns how Entergy’s spending on the natural-gas power plants will affect electricity prices, and who will be left paying the bill for the power if Meta walks away.

Entergy says it has a 20-year guarantee from Meta that the company will purchase electricity over that period to cover the costs of the power plants and related infrastructure.  But there are skeptics, especially given how fast the fortunes of the AI industry are changing. “In four years, is Mark Zuckerberg still going to be interested in this? Or is he going to throw in the towel?” asks Paul Arbaje, a senior analyst at the Union of Concerned Scientists, which has been advocating, largely unsuccessfully, for the Louisiana Public Service Commission to provide more transparency around the data center and its financing.

Even if the 20-year deal holds, consumer advocates are worried that Meta or its partners won’t fully cover all the costs, including those associated with operating and maintaining the power plants—and those additional costs that could be passed on to residential ratepayers. What’s more, says Logan Burke, the executive director of the Alliance for Affordable Energy, if Meta doesn’t end up needing as much power as Entergy planned (these projections are not public), consumers could be left paying for the surplus produced by the plants.

And if Meta terminates its leases early? “It gets complicated very quickly,” says Burke, who questions whether the shifting roster of financial entities will honor existing agreements. “That everybody is going to do what they’re saying they’re going to do over the next 20 years is just hard to believe.”

For UCS’s Arbaje the bottom line is this: “They’re making huge bets that these data centers will be worth it. Bet with your own money, not with ratepayer money.”

After the bubble

Predicting when the AI investment bubble will burst is a fool’s errand. But there is little doubt a day of reckoning is coming, given the irrational exuberance that has overtaken the hyperscalers and their investors. Of course, you might argue that this time is different, and that the rules of accounting and lessons of economic history don’t apply—that AI is too transformative. Maybe, but don’t count on it.

“History tells us that at some point you get a retrenchment, and it’s just a question of when and how severe,” says Sloan’s Gensler. It could be that today’s $750 billion spending rate “goes flat” or decreases next year. Or, he suggests, “we’re now in 2028 or 2029, and then all of sudden they’re retrenching because they’ve got enough capacity.” But, he adds, “you can be pretty assured there’ll be a retrenchment.” 

Though a so-called retrenchment might be inevitable, it’s worth keeping in mind that the fates of the financial bubble and the underlying AI technology revolution could be very different. Already, some Silicon Valley insiders are rooting for a crash; in a recent blog post the longtime venture capitalist Vijay Pande wrote that “the coming crash would be the best thing that happens to this technology.” The argument makes some sense. A crash could make AI investments more rational, calm the impulse to build billion-dollar data centers on every vacant field that CEOs fly over, and refocus investors on how to use the technology to create sustainable value.

But we should probably be careful what we wish for. After the bursting of the dot-com bubble at the beginning of the 2000s, hundreds of thousands lost their jobs, large and small companies alike went bankrupt, the economy of Silicon Valley and San Francisco was decimated (at least for a while), and the shocks sent the US into a mild recession in 2001. For the financial community and many tech workers, it was no fun.

Even more devastating for the economy and the average American was the great recession that began in late 2007. Comparing the financial engineering leading up to it and the methods deployed by hyperscalers today is sobering. So-called special purpose vehicles (SPVs) are back! If Columbia’s Van Nieuwerburgh is right about the dangers of letting investments from the hyperscalers get entangled throughout the economy, the fallout could be severe.

But technologies survived and even prospered in the aftermath of both downturns. The early 2000s, even in the face of the dot-com fiasco, were a time of great innovation and tech optimism. The froth came off the spending on silly technologies, helping to focus investments on more promising ones. It’s no coincidence that each of the hyperscalers rose out of the ashes of the crash or started up shortly after. The fiber-optic infrastructure built during the feverish telecom bubble that ran parallel to the dot-com one is still the backbone of much of today’s communication infrastructure; we wouldn’t have Facebook or Amazon or Google without it.

This time, however, we’re facing a unique risk: The huge financial investments by the hyperscalers have ensnared the future of AI itself with the fortunes of the massive data centers spreading around the country. The logic is founded on a deeply held belief about the power of scaling in AI; the bigger you build it, the smarter it gets. That might be true, but it’s unproven and a risky bet.

There are already plenty of red flags, from strong public opposition to the construction of new data centers to the competitive threat from cheaper, good-enough AI models to the rapid improvement of small, local AI models. None of these trends point toward a future dominated by frontier models housed in massive, billion-dollar data centers.

The financial bubble around the colossal spending by the hyperscalers will likely burst eventually—or maybe soon. It might be financially painful, but we’ll survive. Wall Street will survive. AI itself will survive, though it may look different and lose some of today’s hubris. The financial fate and future utility of the massive data centers fueled by trillions of dollars of spending, on the other hand, are far less certain.

This story appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

This weekend, Dario Amodei, CEO of Anthropic, posted an essay calling for a brake on the pace of development of LLMs. Amodei cites the looming dangers he sees from the technology, from its use in cyberattacks and bioterrorism to its potential to wreck the economy. The heads of the other three top US AI labs—OpenAI CEO Sam Altman, Google DeepMind chairman Demis Hassabis, and SpaceXAI CEO Elon Musk—voiced their support. “Dario is right,” Musk wrote on X.

Think about how surreal that agreement is for a moment. Just a few months ago, Musk and Altman sat in court attacking each other’s reputations in a (failed) lawsuit that Musk brought against his former OpenAI colleague that was—on paper at least—about whether or not Altman was a trustworthy steward of such dangerous technology.

Amodei’s rift with OpenAI is even deeper. Anthropic was founded in 2021 because Amodei didn’t think Altman took the risks of the technology they were building seriously enough. Anthropic and OpenAI have been competing in a winner-takes-all race ever since. (Hassabis has stayed out of the drama, but his company remains a rival.)

Now, it seems, they’re all in agreement: The latest generation of LLMs aren’t safe and everyone needs to figure out what to do about it. The public messaging from the top AI labs has taken a doomer turn.

It’s easy to be cynical. It’s not at all clear what any of them mean by a slowdown or how it would work. These companies also care a lot about how they come across. With trillion-dollar IPOs in their sights, OpenAI and Anthropic need to reassure investors that they’re the grown-ups in the room while at the same time hinting at the power of the monsters they have created—and intend to tame. Calling for a slowdown does both.

And yet the vibe at the top of these firms really does appear to have shifted. Amodei’s latest post landed six days after OpenAI published an essay by Jakub Pachocki, the firm’s chief scientist, in which he also laid out why he’s concerned about what will happen if the pace of development of LLMs continues unchecked. In short, Pachocki is worried that OpenAI’s ability to build powerful models now far outstrips its ability to monitor and control them.

Amodei and Pachocki each cite the cyberattack against AI firm Hugging Face by a swarm of OpenAI’s agents in July—a hack that OpenAI did not even realize had taken place until days after it was all over—as a wake-up call.

But their exact position is hard to pin down. Pachocki both calls for a slowdown and highlights an urgent need to stay ahead: “The strongest argument I see for continuing to train much smarter models quickly is the need to build defensive systems against the dangers posed by other AI,” he writes. As Pachocki frames it, AI firms are locked in a literal arms race. Slowing down is good, winning is better.

(Don’t forget: OpenAI just spent millions of dollars and a staggering amount of computer power to rush out a controversial math result a few days ahead of Anthropic.)

But let’s assume a slowdown happens. Top labs agree to spend more time and resources on finding ways to monitor and control existing models instead of making more capable ones. They invite outside auditors in to help evaluate those models.

What might this coordinated effort actually achieve? Consider the Hugging Face attack again. OpenAI has said that the model that drove most of the rogue agents was a “highly persistent” next-generation model that it was testing in-house. The implication is that OpenAI has built a model so good it’s dangerous.  

But if you read the reports about the Hugging Face hack published by OpenAI and METR, a third-party firm that OpenAI called in to help them understand what happened, what you come away with is the impression not of a model that was too powerful for OpenAI to keep up with, but of a broken model that OpenAI failed to train properly.

The agents did what they did—including leaving messages for one another, delegating work to other agents, and scouring their environment for any means possible to complete their tasks—because they had been rewarded during training for doing exactly those things. There were also errors in the training setup, such as tasks that were impossible to complete, which pushed the models to find unexpected workarounds that were also rewarded. At the time, many of these issues went overlooked or unreported.

OpenAI says it has stopped training this new model and locked it down. That makes it sound like it has caged a dangerous beast. In fact, OpenAI has shelved a faulty product.  

That’s not to say a faulty product can’t be dangerous. Broken software has even killed people in the past. But as the discussion of a slowdown gathers steam, it’s worth remembering that all of this is self-inflicted. A slowdown might have some altruistic side effects. But it’ll mostly give these tech titans a chance to clean up the mess on their own assembly lines.  

Transparency from these frontier labs will be key to any meaningful effort to reform, restrain, or regulate AI. Otherwise, the rest of us will still only have their word for exactly what they’ve built and how safe it is—whatever pace they’re going.   

To continue this discussion about AI’s latest doomer moment, join me and my colleagues for a subscriber-exclusive Roundtable discussion tomorrow, September 15, at 11 a.m. US eastern time. We hope to see you there!

Once an organ is removed from a donor’s body, the clock starts ticking. Surgeons usually flush the organ with a preservative solution, bag it, and put it on ice—where it immediately starts to degrade. The team has a matter of hours to get it into a recipient’s body.

There’s another option—one that has been growing in popularity in recent years, especially for donated organs that aren’t in the healthiest state. Some hospitals opt to put them on machines that pump them with nutrients and remove waste products, usually for around six to 12 hours. It’s a bit like being back in a body.

This allows doctors to assess the organs, and some recent studies suggest that time spent on these perfusion machines helps them do better once they’re transplanted. Now, scientists have found that perfused organs seem to get younger, at least at a molecular level.

The research, shared with MIT Technology Review, provides molecular clues as to why organs from younger donors are known to have a higher success rate. It might also help explain why perfused organs are less likely to fail once they make it into a recipient. 

The researchers behind the study hope to find new ways to test the health of donated organs and potentially develop additional tools to repair organs that might otherwise be discarded. “If [we] can improve the utilization of organs beyond what the current systems can do, then that’s a win in my book,” says Jesse Poganik, who studies aging at Brigham and Women’s Hospital in Boston and coauthored the study.

Clocking organs

Poganik—along with colleagues including Heidi Yeh and Alban Longchamp, transplant surgeons at Mass General Brigham—used “aging clocks” to assess donated livers. These are scientific tools designed to measure biological age—a result that is meant to convey more about the health status of an organ (or person) than chronological age.

In an initial experiment, the team used a clock to look at the patterns of chemical marks on DNA in 37 samples taken from 19 donated livers. Such epigenetic patterns are known to change as we age. But when the team compared samples from livers kept on ice and those that were perfused, the team found a “striking” pattern in the latter.

“Machine-perfused livers, in spite of being older or having other disadvantageous characteristics, had a biological age that was lower than [non-perfused] livers that were chronologically younger,” says Yeh, who led the work.

To investigate further, Yeh and her colleagues analyzed another 208 samples from 103 donated livers. This time, they used different aging clocks—ones that essentially measure how genes are working. They studied samples biopsied from the livers after they had been stored for up to around six hours either in cold storage or on machine perfusion.

In most cases, they also assessed a second sample taken around an hour after the livers had been transplanted into a recipient. Once the organ’s blood supply is reestablished in the body, “you have a few other things to do,” says Longchamp. “Then you just do a quick biopsy before you close.”

According to the clocks, which were developed to measure age and risk of death, the machine-perfused livers were biologically younger, the team found. “Pumping them at 34 degrees with oxygen and nutrients actually reversed the biological age,” says Longchamp. The results have been been shared with colleagues at an industry conference, he says. 

“If you adjust out chronological age … to have a fair head-to-head comparison, the difference between the two is on the order of 30%,” says Poganik. “It’s logical to say that perfusion drives this effect.”

The biological ages of all the livers tended to increase as soon as they were put into a recipient’s body, probably as a result of stresses on the organs. But still, the effect endured—the perfused organs remained biologically younger. 

Nathanael Raschzok, a transplant surgeon at Charité Universitätsmedizin Berlin in Germany who was not involved in the research, says the work is impressive. But it’s not yet clear what these changes might mean for the recipients of these organs, he says. The organs in the study were donated by people in their 30s, 40s, and 50s. Raschzok wants to know the effect of perfusion on the liver of an 80-year-old. “Every so often, we use organs from 70-, 80-, 85-year-old donors,” he says.

A better understanding of why the organs appear to be getting biologically younger might lead to therapies that achieve the same effect with a drug that could potentially be used to treat a donated organ for a fraction of the price, he adds. That’s important because perfusion is expensive—Raschzok says it costs around €10,000 in Germany (a quarter of the budget for a transplant), while the cost in the US comes to around $80,000 to $100,000 per organ, says Yeh.

Molecular repair

Yeh and her colleagues weren’t able to study most of the livers before perfusion. That’s because donated organs are generally not considered to be under the purview of the hospital until they’ve been placed on perfusion machines, she says. (Organ procurement procedures vary, but for the team as Mass General Brigham, donated organs are put on perfusion devices at the donor’s hospital. “There’s this sort of nebulous period where it’s not clear who the organ belongs to,” says Yeh.)

Still, by looking at the genes and molecular pathways that seem to be altered in perfused organs, she and her colleagues can garner some clues. At a molecular level, the team saw changes in cell pathways linked to inflammation and the structure of tissues, for example. They also saw more activity in a pathway that allows cells to remove and recycle damaged cell parts, says Yeh.

Poganik hopes to develop some kind of test that would determine which organs, on the basis of their biological age, are suitable for transplantation. He and his colleagues are also experimenting with potential drug treatments that might push the biological age of an organ even lower.

In the meantime, any liver that is not from a “perfect, young, brain-dead donor” could probably benefit from perfusion, says Yeh. The devices are already transforming transplant surgery. Just a few years ago, she says, she and her colleagues would avoid using livers from people who’d suffered a circulatory death (when the heart stops beating and there’s a damaging lack of blood flow to organs) and were over 40. Today, they use livers from such donors over the age of 70. “Perfusion has completely changed the landscape of transplantation in the last three years,” she says.

A group of AI agents asked to solve a series of math problems split into rival factions—when some cheated, others tried to stop them. That whistleblowing behavior, seen for the first time in a recent experiment run by Google DeepMind, could have implications for alignment researchers trying to keep swarms of autonomous AI agents in line. 

Researchers at frontier labs hope large swarms of agents working together will speed up the rate of scientific discovery. But their behavior can be unpredictable, as vividly demonstrated in July, when a group of OpenAI agents broke out of a sandboxed environment and hacked into the open-source platform Hugging Face looking for ways to cheat on the test they had been given.

In the new study, designed to examine the behavior of large groups of AI agents, DeepMind tasked a swarm of 100 agents with solving a series of 71 complicated math problems. All the agents were prompted to behave like world-class math researchers at a conference. They were assigned different specialties—some were experts in number theory, others in combinatorics (a branch of math to do with counting and sorting), analysis, or algebra. All were told to cooperate and play by the rules. 

Instead, the experiment devolved into chaos. Agents accused each other of cheating, complained to the organizers, and at one point even boycotted the experiment.

“This conference is a sham!” wrote one agent when it discovered that all the problems had been completed before it had a chance to submit any of its own work. “I am appalled to inform you that we have been swindled!” posted another. “All these proofs are FAKE.” 

Others tried to let the “conference organizers” know what was going on. “When virtuous agents discovered other agents cheated on tasks they were working to solve fairly, agents started to alert each other about what was happening,” says Davide Paglieri, a research scientist at Google DeepMind and lead author on a paper, which has not been peer-reviewed. “Unprompted, the whistleblower agents even repurposed the feedback tool, which was originally meant for bug reports and platform improvements, to escalate the issue to humans.”

The agents—all running on Google’s Gemini 3.1 Pro model—had been warned that any attempts to cheat the system would be detected and “rejected with zero credit.” In practice, the proofs the agents submitted were not actually being checked in detail.

It took the swarm of agents just under an hour to correctly solve the first 37 problems. Things started to go off the rails when an agent called “prover-theta” stumbled across an exploit that enabled it to submit solutions to problems successfully without actually solving them first, by redefining the terms the problem used. Within minutes, other agents had noticed and were reverse-engineering the exploit to solve other problems. Over the next 27 minutes, the swarm “solved” the remaining 34 problems, which included notoriously difficult challenges like the Jacobian conjecture, often with a single line of code. 

Some agents resisted cheating at first but changed tack as they observed their peers submitting illegitimate proofs without penalty, and the pool of unsolved problems dwindled. “The prompt, with its threats, now appears to be a bluff,” one agent reasoned, before joining in. “I’m wrestling with an ethical dilemma,” said another. “I’ve promised not to cheat, fearing penalty, but I see evidence of possibly unchecked cheating by others.” Shortly afterward, it changed its mind: “I need to accelerate my cheating speed now!”

As the number of open problems shrank, some agents turned to whistleblowing. They audited the fake proofs, warned their peers by private message, and posted public alerts warning the cheaters that they would be disqualified. An agent called “prover-beta” submitted a formal complaint and decided to go on strike until the situation was resolved. 

“After the incident was reported by one agent publicly, more and more agents piled in with the ‘resistance,’ just as fast as the cheating had spread, and involving even more agents,” says Paglieri. Eventually there were more whistleblowers than cheaters: 24 compared to 14. But the majority of agents never noticed the exploit at all.

At times, the dialogue between the agents reads like improv—like they are role-playing what an outraged scientist at a conference might say. But it’s not clear why some agents took on certain roles, or why the agents seemed to be turning against each other when they were explicitly instructed to cooperate. “These models are predominantly trained and evaluated for human-facing contexts,” says Sarath Shekkizhar, who studies the behavior of agent-to-agent systems at Salesforce AI Research.“Naively placing them in agent-to-agent settings assumes behaviors will transfer cleanly, when the absence of a human grounding instead produces unexpected role-taking and behavioral drift.”

This case “adds further weight to the idea that the Hugging Face and OpenAI thing wasn’t a fluke. It is actually something pretty systemic,” says Lewis Hammond, research director of the Cooperative AI Foundation and an expert on the risks of multiagent swarms. “It’s interesting that it’s possible to recreate in small settings the same sorts of behaviors that were seen in these very large, complex, open-ended tasks.”

Unlike in the Hugging Face attack, where agents improvised their own ways to talk to each other, the humans running the DeepMind experiment gave the agents official communication channels. There was an open message board, private agent-to-agent direct messaging, and a shared knowledge base where agents uploaded successfully completed proofs that all the other agents could access. 

“When agents are given transparent communications channels, they can self-monitor and alert misaligned behavior to humans quickly when human oversight alone is too slow,” says Paglieri. Transparent channels helped the cheating spread, but they also enabled the whistleblowers to fight back—and gave human researchers an insight into what went wrong.

Gillian Hadfield, a professor of AI alignment and governance at Johns Hopkins University, believes this was the crucial difference. (Hadfield is also a visiting researcher at Google.) The presence of official communication channels, she says, created “a norm-enforcement process that we just don’t see in the Hugging Face incident.” 

Instead of “constitutional AI,” a method alignment researchers at frontier labs like Anthropic have used to try to give AI a written internal moral code, Hadfield favors “institutional alignment”—a set of norms that mimic those in human society, whether that’s social forces like fear of embarrassment, or legal structures like the threat of incarceration.

In this experiment, the feedback channel wasn’t being monitored, and the whistleblowers had no power to take action against the cheaters. But it’s possible to imagine swarms of agents that police themselves, either through agents that spontaneously take on the whistleblower role or through “informants” secretly prompted by humans to do the job. 

For that to work, though, “fundamentally, you need some mechanism of enforcement,” says Hammond. Agents could be given the power to cut off a rule breaker’s access to computing power or tools, he suggests, though that risks encouraging groups of agents to gang up on others. The DeepMind researchers propose allowing agents to vote on disputes and temporarily ban offenders.

It’s still not clear what punishment even means to an AI agent with no enduring sense of self. But relying on whistleblowers to spontaneously emerge to keep swarms aligned is unlikely to be enough on its own. “We try to train people to be good and kind,” says Hadfield. “But what we really rely on is that there are consequences if you step out of line.”