Depending who you ask, AI-powered coding is either giving software developers an unprecedented productivity boost or churning out masses of poorly designed code that saps their attention and sets software projects up for serious long term-maintenance problems.

The problem is right now, it’s not easy to know which is true.

As tech giants pour billions into large language models (LLMs), coding has been touted as the technology’s killer app. Both Microsoft CEO Satya Nadella and Google CEO Sundar Pichai have claimed that around a quarter of their companies’ code is now AI-generated. And in March, Anthropic’s CEO, Dario Amodei, predicted that within six months 90% of all code would be written by AI. It’s an appealing and obvious use case. Code is a form of language, we need lots of it, and it’s expensive to produce manually. It’s also easy to tell if it works—run a program and it’s immediately evident whether it’s functional.

---

This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.

---

Executives enamored with the potential to break through human bottlenecks are pushing engineers to lean into an AI-powered future. But after speaking to more than 30 developers, technology executives, analysts, and researchers, MIT Technology Review found that the picture is not as straightforward as it might seem.  

For some developers on the front lines, initial enthusiasm is waning as they bump up against the technology’s limitations. And as a growing body of research suggests that the claimed productivity gains may be illusory, some are questioning whether the emperor is wearing any clothes.

The pace of progress is complicating the picture, though. A steady drumbeat of new model releases mean these tools’ capabilities and quirks are constantly evolving. And their utility often depends on the tasks they are applied to and the organizational structures built around them. All of this leaves developers navigating confusing gaps between expectation and reality. 

Is it the best of times or the worst of times (to channel Dickens) for AI coding? Maybe both.

A fast-moving field

It’s hard to avoid AI coding tools these days. There are a dizzying array of products available, both from model developers like Anthropic, OpenAI, and Google and from companies like Cursor and Windsurf, which wrap these models in polished code-editing software. And according to Stack Overflow’s 2025 Developer Survey, they’re being adopted rapidly, with 65% of developers now using them at least weekly.

AI coding tools first emerged around 2016 but were supercharged with the arrival of LLMs. Early versions functioned as little more than autocomplete for programmers, suggesting what to type next. Today they can analyze entire code bases, edit across files, fix bugs, and even generate documentation explaining how the code works. All this is guided through natural-language prompts via a chat interface.

“Agents”—autonomous LLM-powered coding tools that can take a high-level plan and build entire programs independently—represent the latest frontier in AI coding. This leap was enabled by the latest reasoning models, which can tackle complex problems step by step and, crucially, access external tools to complete tasks. “This is how the model is able to code, as opposed to just talk about coding,” says Boris Cherny, head of Claude Code, Anthropic’s coding agent.

These agents have made impressive progress on software engineering benchmarks—standardized tests that measure model performance. When OpenAI introduced the SWE-bench Verified benchmark in August 2024, offering a way to evaluate agents’ success at fixing real bugs in open-source repositories, the top model solved just 33% of issues. A year later, leading models consistently score above 70%. 

In February, Andrej Karpathy, a founding member of OpenAI and former director of AI at Tesla, coined the term “vibe coding”—meaning an approach where people describe software in natural language and let AI write, refine, and debug the code. Social media abounds with developers who have bought into this vision, claiming massive productivity boosts.

But while some developers and companies report such productivity gains, the hard evidence is more mixed. Early studies from GitHub, Google, and Microsoft—all vendors of AI tools—found developers completing tasks 20% to 55% faster. But a September report from the consultancy Bain & Company described real-world savings as “unremarkable.”

Data from the developer analytics firm GitClear shows that most engineers are producing roughly 10% more durable code—code that isn’t deleted or rewritten within weeks—since 2022, likely thanks to AI. But that gain has come with sharp declines in several measures of code quality. Stack Overflow’s survey also found trust and positive sentiment toward AI tools falling significantly for the first time. And most provocatively, a July study by the nonprofit research organization Model Evaluation & Threat Research (METR) showed that while experienced developers believed AI made them 20% faster, objective tests showed they were actually 19% slower.

Growing disillusionment

For Mike Judge, principal developer at the software consultancy Substantial, the METR study struck a nerve. He was an enthusiastic early adopter of AI tools, but over time he grew frustrated with their limitations and the modest boost they brought to his productivity. “I was complaining to people because I was like, ‘It’s helping me but I can’t figure out how to make it really help me a lot,’” he says. “I kept feeling like the AI was really dumb, but maybe I could trick it into being smart if I found the right magic incantation.”

When asked by a friend, Judge had estimated the tools were providing a roughly 25% speedup. So when he saw similar estimates attributed to developers in the METR study he decided to test his own. For six weeks, he guessed how long a task would take, flipped a coin to decide whether to use AI or code manually, and timed himself. To his surprise, AI slowed him down by an median of 21%—mirroring the METR results.

This got Judge crunching the numbers. If these tools were really speeding developers up, he reasoned, you should see a massive boom in new apps, website registrations, video games, and projects on GitHub. He spent hours and several hundred dollars analyzing all the publicly available data and found flat lines everywhere.

“Shouldn’t this be going up and to the right?” says Judge. “Where’s the hockey stick on any of these graphs? I thought everybody was so extraordinarily productive.” The obvious conclusion, he says, is that AI tools provide little productivity boost for most developers. 

Developers interviewed by MIT Technology Review generally agree on where AI tools excel: producing “boilerplate code” (reusable chunks of code repeated in multiple places with little modification), writing tests, fixing bugs, and explaining unfamiliar code to new developers. Several noted that AI helps overcome the “blank page problem” by offering an imperfect first stab to get a developer’s creative juices flowing. It can also let nontechnical colleagues quickly prototype software features, easing the load on already overworked engineers.

These tasks can be tedious, and developers are typically  glad to hand them off. But they represent only a small part of an experienced engineer’s workload. For the more complex problems where engineers really earn their bread, many developers told MIT Technology Review, the tools face significant hurdles.

Perhaps the biggest problem is that LLMs can hold only a limited amount of information in their “context window”—essentially their working memory. This means they struggle to parse large code bases and are prone to forgetting what they’re doing on longer tasks. “It gets really nearsighted—it’ll only look at the thing that’s right in front of it,” says Judge. “And if you tell it to do a dozen things, it’ll do 11 of them and just forget that last one.”

LLMs’ myopia can lead to headaches for human coders. While an LLM-generated response to a problem may work in isolation, software is made up of hundreds of interconnected modules. If these aren’t built with consideration for other parts of the software, it can quickly lead to a tangled, inconsistent code base that’s hard for humans to parse and, more important, to maintain.

Developers have traditionally addressed this by following conventions—loosely defined coding guidelines that differ widely between projects and teams. “AI has this overwhelming tendency to not understand what the existing conventions are within a repository,” says Bill Harding, the CEO of GitClear. “And so it is very likely to come up with its own slightly different version of how to solve a problem.”

The models also just get things wrong. Like all LLMs, coding models are prone to “hallucinating”—it’s an issue built into how they work. But because the code they output looks so polished, errors can be difficult to detect, says James Liu, director of software engineering at the advertising technology company Mediaocean. Put all these flaws together, and using these tools can feel a lot like pulling a lever on a one-armed bandit. “Some projects you get a 20x improvement in terms of speed or efficiency,” says Liu. “On other things, it just falls flat on its face, and you spend all this time trying to coax it into granting you the wish that you wanted and it’s just not going to.”

Judge suspects this is why engineers often overestimate productivity gains. “You remember the jackpots. You don’t remember sitting there plugging tokens into the slot machine for two hours,” he says.

And it can be particularly pernicious if the developer is unfamiliar with the task. Judge remembers getting AI to help set up a Microsoft cloud service called an Azure Functions, which he’d never used before. He thought it would take about two hours, but nine hours later he threw in the towel. “It kept leading me down these rabbit holes and I didn’t know enough about the topic to be able to tell it ‘Hey, this is nonsensical,’” he says.

The debt begins to mount up

Developers constantly make trade-offs between speed of development and the maintainability of their code—creating what’s known as “technical debt,” says Geoffrey G. Parker, professor of engineering innovation at Dartmouth College. Each shortcut adds complexity and makes the code base harder to manage, accruing “interest” that must eventually be repaid by restructuring the code. As this debt piles up, adding new features and maintaining the software becomes slower and more difficult.

Accumulating technical debt is inevitable in most projects, but AI tools make it much easier for time-pressured engineers to cut corners, says GitClear’s Harding. And GitClear’s data suggests this is happening at scale. Since 2020, the company has seen a significant rise in the amount of copy-pasted code—an indicator that developers are reusing more code snippets, most likely based on AI suggestions—and an even bigger decline in the amount of code moved from one place to another, which happens when developers clean up their code base.

And as models improve, the code they produce is becoming increasingly verbose and complex, says Tariq Shaukat, CEO of Sonar, which makes tools for checking code quality. This is driving down the number of obvious bugs and security vulnerabilities, he says, but at the cost of increasing the number of “code smells”—harder-to-pinpoint flaws that lead to maintenance problems and technical debt. 

Recent research by Sonar found that these make up more than 90% of the issues found in code generated by leading AI models. “Issues that are easy to spot are disappearing, and what’s left are much more complex issues that take a while to find,” says Shaukat. “That’s what worries us about this space at the moment. You’re almost being lulled into a false sense of security.”

If AI tools make it increasingly difficult to maintain code, that could have significant security implications, says Jessica Ji, a security researcher at Georgetown University. “The harder it is to update things and fix things, the more likely a code base or any given chunk of code is to become insecure over time,” says Ji.

There are also more specific security concerns, she says. Researchers have discovered a worrying class of hallucinations where models reference nonexistent software packages in their code. Attackers can exploit this by creating packages with those names that harbor vulnerabilities, which the model or developer may then unwittingly incorporate into software. 

LLMs are also vulnerable to “data-poisoning attacks,” where hackers seed the publicly available data sets models train on with data that alters the model’s behavior in undesirable ways, such as generating insecure code when triggered by specific phrases. In October, research by Anthropic found that as few as 250 malicious documents can introduce this kind of back door into an LLM regardless of its size.

The converted

Despite these issues, though, there’s probably no turning back. “Odds are that writing every line of code on a keyboard by hand—those days are quickly slipping behind us,” says Kyle Daigle, chief operating officer at the Microsoft-owned code-hosting platform GitHub, which produces a popular AI-powered tool called Copilot (not to be confused with the Microsoft product of the same name).

The Stack Overflow report found that despite growing distrust in the technology, usage has increased rapidly and consistently over the past three years. Erin Yepis, a senior analyst at Stack Overflow, says this suggests that engineers are taking advantage of the tools with a clear-eyed view of the risks. The report also found that frequent users tend to be more enthusiastic and more than half of developers are not using the latest coding agents, perhaps explaining why many remain underwhelmed by the technology.

Those latest tools can be a revelation. Trevor Dilley, CTO at the software development agency Twenty20 Ideas, says he had found some value in AI editors’ autocomplete functions, but when he tried anything more complex it would “fail catastrophically.” Then in March, while on vacation with his family, he set the newly released Claude Code to work on one of his hobby projects. It completed a four-hour task in two minutes, and the code was better than what he would have written.

“I was like, Whoa,” he says. “That, for me, was the moment, really. There’s no going back from here.” Dilley has since cofounded a startup called DevSwarm, which is creating software that can marshal multiple agents to work in parallel on a piece of software.

The challenge, says Armin Ronacher, a prominent open-source developer, is that the learning curve for these tools is shallow but long. Until March he’d remained unimpressed by AI tools, but after leaving his job at the software company Sentry in April to launch a startup, he started experimenting with agents. “I basically spent a lot of months doing nothing but this,” he says. “Now, 90% of the code that I write is AI-generated.”

Getting to that point involved extensive trial and error, to figure out which problems tend to trip the tools up and which they can handle efficiently. Today’s models can tackle most coding tasks with the right guardrails, says Ronacher, but these can be very task and project specific.

To get the most out of these tools, developers must surrender control over individual lines of code and focus on the overall software architecture, says Nico Westerdale, chief technology officer at the veterinary staffing company IndeVets. He recently built a data science platform 100,000 lines of code long almost exclusively by prompting models rather than writing the code himself.

Westerdale’s process starts with an extended conversation with the modelagent to develop a detailed plan for what to build and how. He then guides it through each step. It rarely gets things right on the first try and needs constant wrangling, but if you force it to stick to well-defined design patterns, the models can produce high-quality, easily maintainable code, says Westerdale. He reviews every line, and the code is as good as anything he’s ever produced, he says: “I’ve just found it absolutely revolutionary,. It’s also frustrating, difficult, a different way of thinking, and we’re only just getting used to it.”

But while individual developers are learning how to use these tools effectively, getting consistent results across a large engineering team is significantly harder. AI tools amplify both the good and bad aspects of your engineering culture, says Ryan J. Salva, senior director of product management at Google. With strong processes, clear coding patterns, and well-defined best practices, these tools can shine. 

But if your development process is disorganized, they’ll only magnify the problems. It’s also essential to codify that institutional knowledge so the models can draw on it effectively. “A lot of work needs to be done to help build up context and get the tribal knowledge out of our heads,” he says.

The cryptocurrency exchange Coinbase has been vocal about its adoption of AI tools. CEO Brian Armstrong made headlines in August when he revealed that the company had fired staff unwilling to adopt AI tools. But Coinbase’s head of platform, Rob Witoff, tells MIT Technology Review that while they’ve seen massive productivity gains in some areas, the impact has been patchy. For simpler tasks like restructuring the code base and writing tests, AI-powered workflows have achieved speedups of up to 90%. But gains are more modest for other tasks, and the disruption caused by overhauling existing processes often counteracts the increased coding speed, says Witoff.

One factor is that AI tools let junior developers produce far more code,. As in almost all engineering teams, this code has to be reviewed by others, normally more senior developers, to catch bugs and ensure it meets quality standards. But the sheer volume of code now being churned out i whichs quickly saturatinges the ability of midlevel staff to review changes. “This is the cycle we’re going through almost every month, where we automate a new thing lower down in the stack, which brings more pressure higher up in the stack,” he says. “Then we’re looking at applying automation to that higher-up piece.”

Developers also spend only 20% to 40% of their time coding, says Jue Wang, a partner at Bain, so even a significant speedup there often translates to more modest overall gains. Developers spend the rest of their time analyzing software problems and dealing with customer feedback, product strategy, and administrative tasks. To get significant efficiency boosts, companies may need to apply generative AI to all these other processes too, says Jue, and that is still in the works.

Rapid evolution

Programming with agents is a dramatic departure from previous working practices, though, so it’s not surprising companies are facing some teething issues. These are also very new products that are changing by the day. “Every couple months the model improves, and there’s a big step change in the model’s coding capabilities and you have to get recalibrated,” says Anthropic’s Cherny.

For example, in June Anthropic introduced a built-in planning mode to Claude; it has since been replicated by other providers. In October, the company also enabled Claude to ask users questions when it needs more context or faces multiple possible solutions, which Cherny says helps it avoid the tendency to simply assume which path is the best way forward.

Most significant, Anthropic has added features that make Claude better at managing its own context. When it nears the limits of its working memory, it summarizes key details and uses them to start a new context window, effectively giving it an “infinite” one, says Cherny. Claude can also invoke sub-agents to work on smaller tasks, so it no longer has to hold all aspects of the project in its own head. The company claims that its latest model, Claude 4.5 Sonnet, can now code autonomously for more than 30 hours without major performance degradation.

Novel approaches to software development could also sidestep coding agents’ other flaws. MIT professor Max Tegmark has introduced something he calls “vericoding,” which could allow agents to produce entirely bug-free code from a natural-language description. It builds on an approach known as “formal verification,” where developers create a mathematical model of their software that can prove incontrovertibly that it functions correctly. This approach is used in high-stakes areas like flight-control systems and cryptographic libraries, but it remains costly and time-consuming, limiting its broader use.

Rapid improvements in LLMs’ mathematical capabilities have opened up the tantalizing possibility of models that produce not only software but the mathematical proof that it’s bug free, says Tegmark. “You just give the specification, and the AI comes back with provably correct code,” he says. “You don’t have to touch the code. You don’t even have to ever look at the code.”

When tested on about 2,000 vericoding problems in Dafny—a language designed for formal verification—the best LLMs solved over 60%, according to non-peer-reviewed research by Tegmark’s group. This was achieved with off-the-shelf LLMs, and Tegmark expects that training specifically for vericoding could improve scores rapidly.

And counterintuitively, Tthe speed at which AI generates code could actuallylso ease maintainability concerns. Alex Worden, principal engineer at the business software giant Intuit, notes that maintenance is often difficult because engineers reuse components across projects, creating a tangle of dependencies where one change triggers cascading effects across the code base. Reusing code used to save developers time, but in a world where AI can produce hundreds of lines of code in seconds, that imperative has gone, says Worden.

Instead, he advocates for “disposable code,” where each component is generated independently by AI without regard for whether it follows design patterns or conventions. They are then connected via APIs—sets of rules that let components request information or services from each other. Each component’s inner workings are not dependent on other parts of the code base, making it possible to rip them out and replace them without wider impact, says Worden. 

“The industry is still concerned about humans maintaining AI-generated code,” he says. “I question how long humans will look at or care about code.”

A narrowing talent pipeline

For the foreseeable future, though, humans will still need to understand and maintain the code that underpins their projects. And one of the most pernicious side effects of AI tools may be a shrinking pool of people capable of doing so. 

Early evidence suggests that fears around the job-destroying effects of AI may be justified. A recent Stanford University study found that employment among software developers aged 22 to 25 fell nearly 20% between 2022 and 2025, coinciding with the rise of AI-powered coding tools.

Experienced developers could face difficulties too. Luciano Nooijen, an engineer at the video-game infrastructure developer Companion Group, used AI tools heavily in his day job, where they were provided for free. But when he began a side project without access to those tools, he found himself struggling with tasks that previously came naturally. “I was feeling so stupid because things that used to be instinct became manual, sometimes even cumbersome,” says Nooijen.

Just as athletes still perform basic drills, he thinks the only way to maintain an instinct for coding is to regularly practice the grunt work. That’s why he’s largely abandoned AI tools, though he admits that deeper motivations are also at play. 

Part of the reason Nooijen and other developers MIT Technology Review spoke to are pushing back against AI tools is a sense that they are hollowing out the parts of their jobs that they love. “I got into software engineering because I like working with computers. I like making machines do things that I want,” Nooijen says. “It’s just not fun sitting there with my work being done for me.”

Each time you’ve heard a borderline outlandish idea of what AI will be capable of, it often turns out that Sam Altman was, if not the first to articulate it, at least the most persuasive and influential voice behind it. 

For more than a decade he has been known in Silicon Valley as a world-class fundraiser and persuader. OpenAI’s early releases around 2020 set the stage for a mania around large language models, and the launch of ChatGPT in November 2022 granted Altman a world stage on which to present his new thesis: that these models mirror human intelligence and could swing the doors open to a healthier and wealthier techno-utopia.

---

This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.

---

Throughout, Altman’s words have set the agenda. He has framed a prospective superintelligent AI as either humanistic or catastrophic, depending on what effect he was hoping to create, what he was raising money for, or which tech giant seemed like his most formidable competitor at the moment. 

Examining Altman’s statements over the years reveals just how much his outlook has powered today’s AI boom. Even among Silicon Valley’s many hypesters, he’s been especially willing to speak about open questions—whether large language models contain the ingredients of human thought, whether language can also produce intelligence—as if they were already answered. 

In conclusion … 

Altman didn’t dupe the world. OpenAI has ushered in a genuine tech revolution, with increasingly impressive language models that have attracted millions of users. Even skeptics would concede that LLMs’ conversational ability is astonishing.

But Altman’s hype has always hinged less on today’s capabilities than on a philosophical tomorrow—an outlook that quite handily doubles as a case for more capital and friendlier regulation. Long before large language models existed, he was imagining an AI powerful enough to require wealth redistribution, just as he imagined humanity colonizing other planets. Again and again, promises of a destination—abundance, superintelligence, a healthier and wealthier world—have come first, and the evidence second. 

Even if LLMs eventually hit a wall, there’s little reason to think his faith in a techno-utopian future will falter. The vision was never really about the particulars of the current model anyway. 

It’s a weird time to be an AI doomer.

This small but influential community of researchers, scientists, and policy experts believes, in the simplest terms, that AI could get so good it could be bad—very, very bad—for humanity. Though many of these people would be more likely to describe themselves as advocates for AI safety than as literal doomsayers, they warn that AI poses an existential risk to humanity. They argue that absent more regulation, the industry could hurtle toward systems it can’t control. They commonly expect such systems to follow the creation of artificial general intelligence (AGI), a slippery concept generally understood as technology that can do whatever humans can do, and better. 

---

This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.

---

Though this is far from a universally shared perspective in the AI field, the doomer crowd has had some notable success over the past several years: helping shape AI policy coming from the Biden administration, organizing prominent calls for international “red lines” to prevent AI risks, and getting a bigger (and more influential) megaphone as some of its adherents win science’s most prestigious awards.

But a number of developments over the past six months have put them on the back foot. Talk of an AI bubble has overwhelmed the discourse as tech companies continue to invest in multiple Manhattan Projects’ worth of data centers without any certainty that future demand will match what they’re building. 

And then there was the August release of OpenAI’s latest foundation model, GPT-5, which proved something of a letdown. Maybe that was inevitable, since it was the most hyped AI release of all time; OpenAI CEO Sam Altman had boasted that GPT-5 felt “like a PhD-level expert” in every topic and told the podcaster Theo Von that the model was so good, it had made him feel “useless relative to the AI.” 

Many expected GPT-5 to be a big step toward AGI, but whatever progress the model may have made was overshadowed by a string of technical bugs and the company’s mystifying, quickly reversed decision to shut off access to every old OpenAI model without warning. And while the new model achieved state-of-the-art benchmark scores, many people felt, perhaps unfairly, that in day-to-day use GPT-5 was a step backward. 

All this would seem to threaten some of the very foundations of the doomers’ case. In turn, a competing camp of AI accelerationists, who fear AI is actually not moving fast enough and that the industry is constantly at risk of being smothered by overregulation, is seeing a fresh chance to change how we approach AI safety (or, maybe more accurately, how we don’t). 

This is particularly true of the industry types who’ve decamped to Washington: “The Doomer narratives were wrong,” declared David Sacks, the longtime venture capitalist turned Trump administration AI czar. “This notion of imminent AGI has been a distraction and harmful and now effectively proven wrong,” echoed the White House’s senior policy advisor for AI and tech investor Sriram Krishnan. (Sacks and Krishnan did not reply to requests for comment.) 

(There is, of course, another camp in the AI safety debate: the group of researchers and advocates commonly associated with the label “AI ethics.” Though they also favor regulation, they tend to think the speed of AI progress has been overstated and have often written off AGI as a sci-fi story or a scam that distracts us from the technology’s immediate threats. But any potential doomer demise wouldn’t exactly give them the same opening the accelerationists are seeing.)

So where does this leave the doomers? As part of our Hype Correction package, we decided to ask some of the movement’s biggest names to see if the recent setbacks and general vibe shift had altered their views. Are they frustrated that policymakers no longer seem to heed their threats? Are they quietly adjusting their timelines for the apocalypse? 

Recent interviews with 20 people who study or advocate AI safety and governance—including Nobel Prize winner Geoffrey Hinton, Turing Prize winner Yoshua Bengio, and high-profile experts like former OpenAI board member Helen Toner—reveal that rather than feeling chastened or lost in the wilderness, they’re still deeply committed to their cause, believing that AGI remains not just possible but incredibly dangerous.

At the same time, they seem to be grappling with a near contradiction. While they’re somewhat relieved that recent developments suggest AGI is further out than they previously thought (“Thank God we have more time,” says AI researcher Jeffrey Ladish), they also feel angry that people in power are not taking them seriously enough (Daniel Kokotajlo, lead author of a cautionary forecast called “AI 2027,” calls the Sacks and Krishnan tweets “deranged and/or dishonest”). 

Broadly speaking, these experts see the talk of an AI bubble as no more than a speed bump, and disappointment in GPT-5 as more distracting than illuminating. They still generally favor more robust regulation and worry that progress on policy—the implementation of the EU AI Act; the passage of the first major American AI safety bill, California’s SB 53; and new interest in AGI risk from some members of Congress—has become vulnerable as Washington overreacts to what doomers see as short-term failures to live up to the hype. 

Some were also eager to correct what they see as the most persistent misconceptions about the doomer world. Though their critics routinely mock them for predicting that AGI is right around the corner, they claim that’s never been an essential part of their case: It “isn’t about imminence,” says Berkeley professor Stuart Russell, the author of Human Compatible: Artificial Intelligence and the Problem of Control. Most people I spoke with say their timelines to dangerous systems have actually lengthened slightly in the last year—an important change given how quickly the policy and technical landscapes can shift. 

“If someone said there’s a four-mile-diameter asteroid that’s going to hit the Earth in 2067, we wouldn’t say, ‘Remind me in 2066 and we’ll think about it.’”

Many of them, in fact, emphasize the importance of changing timelines. And even if they are just a tad longer now, Toner tells me that one big-picture story of the ChatGPT era is the dramatic compression of these estimates across the AI world. For a long while, she says, AGI was expected in many decades. Now, for the most part, the predicted arrival is sometime in the next few years to 20 years. So even if we have a little bit more time, she (and many of her peers) continue to see AI safety as incredibly, vitally urgent. She tells me that if AGI were possible anytime in even the next 30 years, “It’s a huge fucking deal. We should have a lot of people working on this.”

So despite the precarious moment doomers find themselves in, their bottom line remains that no matter when AGI is coming (and, again, they say it’s very likely coming), the world is far from ready. 

Maybe you agree. Or maybe you may think this future is far from guaranteed. Or that it’s the stuff of science fiction. You may even think AGI is a great big conspiracy theory. You’re not alone, of course—this topic is polarizing. But whatever you think about the doomer mindset, there’s no getting around the fact that certain people in this world have a lot of influence. So here are some of the most prominent people in the space, reflecting on this moment in their own words. 

Interviews have been edited and condensed for length and clarity. 

---

The Nobel laureate who’s not sure what’s coming

Geoffrey Hinton, winner of the Turing Award and the Nobel Prize in physics for pioneering deep learning

The biggest change in the last few years is that there are people who are hard to dismiss who are saying this stuff is dangerous. Like, [former Google CEO] Eric Schmidt, for example, really recognized this stuff could be really dangerous. He and I were in China recently talking to someone on the Politburo, the party secretary of Shanghai, to make sure he really understood—and he did. I think in China, the leadership understands AI and its dangers much better because many of them are engineers.

I’ve been focused on the longer-term threat: When AIs get more intelligent than us, can we really expect that humans will remain in control or even relevant? But I don’t think anything is inevitable. There’s huge uncertainty on everything. We’ve never been here before. Anybody who’s confident they know what’s going to happen seems silly to me. I think this is very unlikely but maybe it’ll turn out that all the people saying AI is way overhyped are correct. Maybe it’ll turn out that we can’t get much further than the current chatbots—we hit a wall due to limited data. I don’t believe that. I think that’s unlikely, but it’s possible. 

I also don’t believe people like Eliezer Yudkowsky, who say if anybody builds it, we’re all going to die. We don’t know that. 

But if you go on the balance of the evidence, I think it’s fair to say that most experts who know a lot about AI believe it’s very probable that we’ll have superintelligence within the next 20 years. [Google DeepMind CEO] Demis Hassabis says maybe 10 years. Even [prominent AI skeptic] Gary Marcus would probably say, “Well, if you guys make a hybrid system with good old-fashioned symbolic logic … maybe that’ll be superintelligent.” [Editor’s note: In September, Marcus predicted AGI would arrive between 2033 and 2040.]

And I don’t think anybody believes progress will stall at AGI. I think more or less everybody believes a few years after AGI, we’ll have superintelligence, because the AGI will be better than us at building AI.

So while I think it’s clear that the winds are getting more difficult, simultaneously, people are putting in many more resources [into developing advanced AI]. I think progress will continue just because there’s many more resources going in.

The deep learning pioneer who wishes he’d seen the risks sooner

Yoshua Bengio, winner of the Turing Award, chair of the International AI Safety Report, and founder of LawZero

Some people thought that GPT-5 meant we had hit a wall, but that isn’t quite what you see in the scientific data and trends.

There have been people overselling the idea that AGI is tomorrow morning, which commercially could make sense. But if you look at the various benchmarks, GPT-5 is just where you would expect the models at that point in time to be. By the way, it’s not just GPT-5, it’s Claude and Google models, too. In some areas where AI systems weren’t very good, like Humanity’s Last Exam or FrontierMath, they’re getting much better scores now than they were at the beginning of the year.

At the same time, the overall landscape for AI governance and safety is not good. There’s a strong force pushing against regulation. It’s like climate change. We can put our head in the sand and hope it’s going to be fine, but it doesn’t really deal with the issue.

The biggest disconnect with policymakers is a misunderstanding of the scale of change that is likely to happen if the trend of AI progress continues. A lot of people in business and governments simply think of AI as just another technology that’s going to be economically very powerful. They don’t understand how much it might change the world if trends continue, and we approach human-level AI. 

Like many people, I had been blinding myself to the potential risks to some extent. I should have seen it coming much earlier. But it’s human. You’re excited about your work and you want to see the good side of it. That makes us a little bit biased in not really paying attention to the bad things that could happen.

Even a small chance—like 1% or 0.1%—of creating an accident where billions of people die is not acceptable. 

The AI veteran who believes AI is progressing—but not fast enough to prevent the bubble from bursting

Stuart Russell, distinguished professor of computer science, University of California, Berkeley, and author of Human Compatible

I hope the idea that talking about existential risk makes you a “doomer” or is “science fiction” comes to be seen as fringe, given that most leading AI researchers and most leading AI CEOs take it seriously. 

There have been claims that AI could never pass a Turing test, or you could never have a system that uses natural language fluently, or one that could parallel-park a car. All these claims just end up getting disproved by progress.

People are spending trillions of dollars to make superhuman AI happen. I think they need some new ideas, but there’s a significant chance they will come up with them, because many significant new ideas have happened in the last few years. 

My fairly consistent estimate for the last 12 months has been that there’s a 75% chance that those breakthroughs are not going to happen in time to rescue the industry from the bursting of the bubble. Because the investments are consistent with a prediction that we’re going to have much better AI that will deliver much more value to real customers. But if those predictions don’t come true, then there’ll be a lot of blood on the floor in the stock markets.

However, the safety case isn’t about imminence. It’s about the fact that we still don’t have a solution to the control problem. If someone said there’s a four-mile-diameter asteroid that’s going to hit the Earth in 2067, we wouldn’t say, “Remind me in 2066 and we’ll think about it.” We don’t know how long it takes to develop the technology needed to control superintelligent AI.

Looking at precedents, the acceptable level of risk for a nuclear plant melting down is about one in a million per year. Extinction is much worse than that. So maybe set the acceptable risk at one in a billion. But the companies are saying it’s something like one in five. They don’t know how to make it acceptable. And that’s a problem.

The professor trying to set the narrative straight on AI safety

David Krueger, assistant professor in machine learning at the University of Montreal and Yoshua Bengio’s Mila Institute, and founder of Evitable

I think people definitely overcorrected in their response to GPT-5. But there was hype. My recollection was that there were multiple statements from CEOs at various levels of explicitness who basically said that by the end of 2025, we’re going to have an automated drop-in replacement remote worker. But it seems like it’s been underwhelming, with agents just not really being there yet.

I’ve been surprised how much these narratives predicting AGI in 2027 capture the public attention. When 2027 comes around, if things still look pretty normal, I think people are going to feel like the whole worldview has been falsified. And it’s really annoying how often when I’m talking to people about AI safety, they assume that I think we have really short timelines to dangerous systems, or that I think LLMs or deep learning are going to give us AGI. They ascribe all these extra assumptions to me that aren’t necessary to make the case. 

I’d expect we need decades for the international coordination problem. So even if dangerous AI is decades off, it’s already urgent. That point seems really lost on a lot of people. There’s this idea of “Let’s wait until we have a really dangerous system and then start governing it.” Man, that is way too late.

I still think people in the safety community tend to work behind the scenes, with people in power, not really with civil society. It gives ammunition to people who say it’s all just a scam or insider lobbying. That’s not to say that there’s no truth to these narratives, but the underlying risk is still real. We need more public awareness and a broad base of support to have an effective response.

If you actually believe there’s a 10% chance of doom in the next 10 years—which I think a reasonable person should, if they take a close look—then the first thing you think is: “Why are we doing this? This is crazy.” That’s just a very reasonable response once you buy the premise.

The governance expert worried about AI safety’s credibility

Helen Toner, acting executive director of Georgetown University’s Center for Security and Emerging Technology and former OpenAI board member

When I got into the space, AI safety was more of a set of philosophical ideas. Today, it’s a thriving set of subfields of machine learning, filling in the gulf between some of the more “out there” concerns about AI scheming, deception, or power-seeking and real concrete systems we can test and play with. 

“I worry that some aggressive AGI timeline estimates from some AI safety people are setting them up for a boy-who-cried-wolf moment.”

AI governance is improving slowly. If we have lots of time to adapt and governance can keep improving slowly, I feel not bad. If we don’t have much time, then we’re probably moving too slow.

I think GPT-5 is generally seen as a disappointment in DC. There’s a pretty polarized conversation around: Are we going to have AGI and superintelligence in the next few years? Or is AI actually just totally all hype and useless and a bubble? The pendulum had maybe swung too far toward “We’re going to have super-capable systems very, very soon.” And so now it’s swinging back toward “It’s all hype.”

I worry that some aggressive AGI timeline estimates from some AI safety people are setting them up for a boy-who-cried-wolf moment. When the predictions about AGI coming in 2027 don’t come true, people will say, “Look at all these people who made fools of themselves. You should never listen to them again.” That’s not the intellectually honest response, if maybe they later changed their mind, or their take was that they only thought it was 20 percent likely and they thought that was still worth paying attention to. I think that shouldn’t be disqualifying for people to listen to you later, but I do worry it will be a big credibility hit. And that’s applying to people who are very concerned about AI safety and never said anything about very short timelines.

The AI security researcher who now believes AGI is further out—and is grateful

Jeffrey Ladish, executive director at Palisade Research

In the last year, two big things updated my AGI timelines. 

First, the lack of high-quality data turned out to be a bigger problem than I expected. 

Second, the first “reasoning” model, OpenAI’s o1 in September 2024, showed reinforcement learning scaling was more effective than I thought it would be. And then months later, you see the o1 to o3 scale-up and you see pretty crazy impressive performance in math and coding and science—domains where it’s easier to sort of verify the results. But while we’re seeing continued progress, it could have been much faster.

All of this bumps up my median estimate to the start of fully automated AI research and development from three years to maybe five or six years. But those are kind of made up numbers. It’s hard. I want to caveat all this with, like, “Man, it’s just really hard to do forecasting here.”

Thank God we have more time. We have a possibly very brief window of opportunity to really try to understand these systems before they are capable and strategic enough to pose a real threat to our ability to control them.

But it’s scary to see people think that we’re not making progress anymore when that’s clearly not true. I just know it’s not true because I use the models. One of the downsides of the way AI is progressing is that how fast it’s moving is becoming less legible to normal people. 

Now, this is not true in some domains—like, look at Sora 2. It is so obvious to anyone who looks at it that Sora 2 is vastly better than what came before. But if you ask GPT-4 and GPT-5 why the sky is blue, they’ll give you basically the same answer. It is the correct answer. It’s already saturated the ability to tell you why the sky is blue. So the people who I expect to most understand AI progress right now are the people who are actually building with AIs or using AIs on very difficult scientific problems.

The AGI forecaster who saw the critics coming

Daniel Kokotajlo, executive director of the AI Futures Project; an OpenAI whistleblower; and lead author of “AI 2027,” a vivid scenario where—starting in 2027—AIs progress from “superhuman coders” to “wildly superintelligent” systems in the span of months

AI policy seems to be getting worse, like the “Pro-AI” super PAC [launched earlier this year by executives from OpenAI and Andreessen Horowitz to lobby for a deregulatory agenda], and the deranged and/or dishonest tweets from Sriram Krishnan and David Sacks. AI safety research is progressing at the usual pace, which is excitingly rapid compared to most fields, but slow compared to how fast it needs to be.

We said on the first page of “AI 2027” that our timelines were somewhat longer than 2027. So even when we launched AI 2027, we expected there to be a bunch of critics in 2028 triumphantly saying we’ve been discredited, like the tweets from Sacks and Krishnan. But we thought, and continue to think, that the intelligence explosion will probably happen sometime in the next five to 10 years, and that when it does, people will remember our scenario and realize it was closer to the truth than anything else available in 2025. 

Predicting the future is hard, but it’s valuable to try; people should aim to communicate their uncertainty about the future in a way that is specific and falsifiable. This is what we’ve done and very few others have done. Our critics mostly haven’t made predictions of their own and often exaggerate and mischaracterize our views. They say our timelines are shorter than they are or ever were, or they say we are more confident than we are or were.

I feel pretty good about having longer timelines to AGI. It feels like I just got a better prognosis from my doctor. The situation is still basically the same, though.

Garrison Lovely is a freelance journalist and the author of Obsolete, an online publication and forthcoming book on the discourse, economics, and geopolitics of the race to build machine superintelligence (out spring 2026). His writing on AI has appeared in the New York Times, Nature, Bloomberg, Time, the Guardian, The Verge, and elsewhere.

Some disillusionment was inevitable. When OpenAI released a free web app called ChatGPT in late 2022, it changed the course of an entire industry—and several world economies. Millions of people started talking to their computers, and their computers started talking back. We were enchanted, and we expected more.

We got it. Technology companies scrambled to stay ahead, putting out rival products that outdid one another with each new release: voice, images, video. With nonstop one-upmanship, AI companies have presented each new product drop as a major breakthrough, reinforcing a widespread faith that this technology would just keep getting better. Boosters told us that progress was exponential. They posted charts plotting how far we’d come since last year’s models: Look how the line goes up! Generative AI could do anything, it seemed.

Well, 2025 has been a year of reckoning. 

---

This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.

---

For a start, the heads of the top AI companies made promises they couldn’t keep. They told us that generative AI would replace the white-collar workforce, bring about an age of abundance, make scientific discoveries, and help find new cures for disease. FOMO across the world’s economies, at least in the Global North, made CEOs tear up their playbooks and try to get in on the action.

That’s when the shine started to come off. Though the technology may have been billed as a universal multitool that could revamp outdated business processes and cut costs, a number of studies published this year suggest that firms are failing to make the AI pixie dust work its magic. Surveys and trackers from a range of sources, including the US Census Bureau and Stanford University, have found that business uptake of AI tools is stalling. And when the tools do get tried out, many projects stay stuck in the pilot stage. Without broad buy-in across the economy it is not clear how the big AI companies will ever recoup the incredible amounts they’ve already spent in this race. 

At the same time, updates to the core technology are no longer the step changes they once were.

The highest-profile example of this was the botched launch of GPT-5 in August. Here was OpenAI, the firm that had ignited (and to a large extent sustained) the current boom, set to release a brand-new generation of its technology. OpenAI had been hyping GPT-5 for months: “PhD-level expert in anything,” CEO Sam Altman crowed. On another occasion Altman posted, without comment, an image of the Death Star from Star Wars, which OpenAI stans took to be a symbol of ultimate power: Coming soon! Expectations were huge.

And yet, when it landed, GPT-5 seemed to be—more of the same? What followed was the biggest vibe shift since ChatGPT first appeared three years ago. “The era of boundary-breaking advancements is over,” Yannic Kilcher, an AI researcher and popular YouTuber, announced in a video posted two days after GPT-5 came out: “AGI is not coming. It seems very much that we’re in the Samsung Galaxy era of LLMs.”

A lot of people (me included) have made the analogy with phones. For a decade or so, smartphones were the most exciting consumer tech in the world. Today, new products drop from Apple or Samsung with little fanfare. While superfans pore over small upgrades, to most people this year’s iPhone now looks and feels a lot like last year’s iPhone. Is that where we are with generative AI? And is it a problem? Sure, smartphones have become the new normal. But they changed the way the world works, too.

To be clear, the last few years have been filled with genuine “Wow” moments, from the stunning leaps in the quality of video generation models to the problem-solving chops of so-called reasoning models to the world-class competition wins of the latest coding and math models. But this remarkable technology is only a few years old, and in many ways it is still experimental. Its successes come with big caveats.

Perhaps we need to readjust our expectations.

The big reset

Let’s be careful here: The pendulum from hype to anti-hype can swing too far. It would be rash to dismiss this technology just because it has been oversold. The knee-jerk response when AI fails to live up to its hype is to say that progress has hit a wall. But that misunderstands how research and innovation in tech work. Progress has always moved in fits and starts. There are ways over, around, and under walls.

Take a step back from the GPT-5 launch. It came hot on the heels of a series of remarkable models that OpenAI had shipped in the previous months, including o1 and o3 (first-of-their-kind reasoning models that introduced the industry to a whole new paradigm) and Sora 2, which raised the bar for video generation once again. That doesn’t sound like hitting a wall to me.

AI is really good! Look at Nano Banana Pro, the new image generation model from Google DeepMind that can turn a book chapter into an infographic, and much more. It’s just there—for free—on your phone.

And yet you can’t help but wonder: When the wow factor is gone, what’s left? How will we view this technology a year or five from now? Will we think it was worth the colossal costs, both financial and environmental? 

With that in mind, here are four ways to think about the state of AI at the end of 2025: The start of a much-needed hype correction.

01: LLMs are not everything

In some ways, it is the hype around large language models, not AI as a whole, that needs correcting. It has become obvious that LLMs are not the doorway to artificial general intelligence, or AGI, a hypothetical technology that some insist will one day be able to do any (cognitive) task a human can.

Even an AGI evangelist like Ilya Sutskever, chief scientist and cofounder at the AI startup Safe Superintelligence and former chief scientist and cofounder at OpenAI, now highlights the limitations of LLMs, a technology he had a huge hand in creating. LLMs are very good at learning how to do a lot of specific tasks, but they do not seem to learn the principles behind those tasks, Sutskever said in an interview with Dwarkesh Patel in November.

It’s the difference between learning how to solve a thousand different algebra problems and learning how to solve any algebra problem. “The thing which I think is the most fundamental is that these models somehow just generalize dramatically worse than people,” Sutskever said.

It’s easy to imagine that LLMs can do anything because their use of language is so compelling. It is astonishing how well this technology can mimic the way people write and speak. And we are hardwired to see intelligence in things that behave in certain ways—whether it’s there or not. In other words, we have built machines with humanlike behavior and cannot resist seeing a humanlike mind behind them.

That’s understandable. LLMs have been part of mainstream life for only a few years. But in that time, marketers have preyed on our shaky sense of what the technology can really do, pumping up expectations and turbocharging the hype. As we live with this technology and come to understand it better, those expectations should fall back down to earth.  

02: AI is not a quick fix to all your problems

In July, researchers at MIT published a study that became a tentpole talking point in the disillusionment camp. The headline result was that a whopping 95% of businesses that had tried using AI had found zero value in it.  

The general thrust of that claim was echoed by other research, too. In November, a study by researchers at Upwork, a company that runs an online marketplace for freelancers, found that agents powered by top LLMs from OpenAI, Google DeepMind, and Anthropic failed to complete many straightforward workplace tasks by themselves.

This is miles off Altman’s prediction: “We believe that, in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies,” he wrote on his personal blog in January.

But what gets missed in that MIT study is that the researchers’ measure of success was pretty narrow. That 95% failure rate accounts for companies that had tried to implement bespoke AI systems but had not yet scaled them beyond the pilot stage after six months. It shouldn’t be too surprising that a lot of experiments with experimental technology don’t pan out straight away.

That number also does not include the use of LLMs by employees outside of official pilots. The MIT researchers found that around 90% of the companies they surveyed had a kind of AI shadow economy where workers were using personal chatbot accounts. But the value of that shadow economy was not measured.  

When the Upwork study looked at how well agents completed tasks together with people who knew what they were doing, success rates shot up. The takeaway seems to be that a lot of people are figuring out for themselves how AI might help them with their jobs.

That fits with something the AI researcher and influencer (and coiner of the term “vibe coding”) Andrej Karpathy has noted: Chatbots are better than the average human at a lot of different things (think of giving legal advice, fixing bugs, doing high school math), but they are not better than an expert human. Karpathy suggests this may be why chatbots have proved popular with individual consumers, helping non-experts with everyday questions and tasks, but they have not upended the economy, which would require outperforming skilled employees at their jobs.

That may change. For now, don’t be surprised that AI has not (yet) had the impact on jobs that boosters said it would. AI is not a quick fix, and it cannot replace humans. But there’s a lot to play for. The ways in which AI could be integrated into everyday workflows and business pipelines are still being tried out.   

03: Are we in a bubble? (If so, what kind of bubble?)

If AI is a bubble, is it like the subprime mortgage bubble of 2008 or the internet bubble of 2000? Because there’s a big difference.

The subprime bubble wiped out a big part of the economy, because when it burst it left nothing behind except debt and overvalued real estate. The dot-com bubble wiped out a lot of companies, which sent ripples across the world, but it left behind the infant internet—an international network of cables and a handful of startups, like Google and Amazon, that became the tech giants of today.  

Then again, maybe we’re in a bubble unlike either of those. After all, there’s no real business model for LLMs right now. We don’t yet know what the killer app will be, or if there will even be one. 

And many economists are concerned about the unprecedented amounts of money being sunk into the infrastructure required to build capacity and serve the projected demand. But what if that demand doesn’t materialize? Add to that the weird circularity of many of those deals—with Nvidia paying OpenAI to pay Nvidia, and so on—and it’s no surprise everybody’s got a different take on what’s coming. 

Some investors remain sanguine. In an interview with the Technology Business Programming Network podcast in November, Glenn Hutchins, cofounder of Silver Lake Partners, a major international private equity firm, gave a few reasons not to worry. “Every one of these data centers—almost all of them—has a solvent counterparty that is contracted to take all the output they’re built to suit,” he said. In other words, it’s not a case of “Build it and they’ll come”—the customers are already locked in. 

And, he pointed out, one of the biggest of those solvent counterparties is Microsoft. “Microsoft has the world’s best credit rating,” Hutchins said. “If you sign a deal with Microsoft to take the output from your data center, Satya is good for it.”

Many CEOs will be looking back at the dot-com bubble and trying to learn its lessons. Here’s one way to see it: The companies that went bust back then didn’t have the money to last the distance. Those that survived the crash thrived.

With that lesson in mind, AI companies today are trying to pay their way through what may or may not be a bubble. Stay in the race; don’t get left behind. Even so, it’s a desperate gamble.

But there’s another lesson too. Companies that might look like sideshows can turn into unicorns fast. Take Synthesia, which makes avatar generation tools for businesses. Nathan Benaich, cofounder of the VC firm Air Street Capital, admits that when he first heard about the company a few years ago, back when fear of deepfakes was rife, he wasn’t sure what its tech was for and thought there was no market for it.

“We didn’t know who would pay for lip-synching and voice cloning,” he says. “Turns out there’s a lot of people who wanted to pay for it.” Synthesia now has around 55,000 corporate customers and brings in around $150 million a year. In October, the company was valued at $4 billion.

04: ChatGPT was not the beginning, and it won’t be the end

ChatGPT was the culmination of a decade’s worth of progress in deep learning, the technology that underpins all of modern AI. The seeds of deep learning itself were planted in the 1980s. The field as a whole goes back at least to the 1950s. If progress is measured against that backdrop, generative AI has barely got going.

Meanwhile, research is at a fever pitch. There are more high-quality submissions to the world’s major AI conferences than ever before. This year, organizers of some of those conferences resorted to turning down papers that reviewers had already approved, just to manage numbers. (At the same time, preprint servers like arXiv have been flooded with AI-generated research slop.)

“It’s back to the age of research again,” Sutskever said in that Dwarkesh interview, talking about the current bottleneck with LLMs. That’s not a setback; that’s the start of something new.

“There’s always a lot of hype beasts,” says Benaich. But he thinks there’s an upside to that: Hype attracts the money and talent needed to make real progress. “You know, it was only like two or three years ago that the people who built these models were basically research nerds that just happened on something that kind of worked,” he says. “Now everybody who’s good at anything in technology is working on this.”

Where do we go from here?

The relentless hype hasn’t come just from companies drumming up business for their vastly expensive new technologies. There’s a large cohort of people—inside and outside the industry—who want to believe in the promise of machines that can read, write, and think. It’s a wild decades-old dream. 

But the hype was never sustainable—and that’s a good thing. We now have a chance to reset expectations and see this technology for what it really is—assess its true capabilities, understand its flaws, and take the time to learn how to apply it in valuable (and beneficial) ways. “We’re still trying to figure out how to invoke certain behaviors from this insanely high-dimensional black box of information and skills,” says Benaich.

This hype correction was long overdue. But know that AI isn’t going anywhere. We don’t even fully understand what we’ve built so far, let alone what’s coming next.

On April 28, 2022, at a highly anticipated concert in Spokane, Washington, the musician Paul McCartney astonished his audience with a groundbreaking application of AI: He began to perform with a lifelike depiction of his long-deceased musical partner, John Lennon. 

Using recent advances in audio and video processing, engineers had taken the pair’s final performance (London, 1969), separated Lennon’s voice and image from the original mix and restored them with lifelike clarity.

---

This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.

---

For years, researchers like me had taught machines to “see” and “hear” in order to make such a moment possible. As McCartney and Lennon appeared to reunite across time and space, the arena fell silent; many in the crowd began to cry. As an AI scientist and lifelong Beatles fan, I felt profound gratitude that we could experience this truly life-changing moment. 

Later that year, the world was captivated by another major breakthrough: AI conversation. For the first time in history, systems capable of generating new, contextually relevant comments in real time, on virtually any subject, were widely accessible owing to the release of ChatGPT. Billions of people were suddenly able to interact with AI. This ignited the public’s imagination about what AI could be, bringing an explosion of creative ideas, hopes, and fears.

Having done my PhD on AI language generation (long considered niche), I was thrilled we had come this far. But the awe I felt was rivaled by my growing rage at the flood of media takes and self-appointed experts insisting that generative AI could do things it simply can’t, and warning that anyone who didn’t adopt it would be left behind.

This kind of hype has contributed to a frenzy of misunderstandings about what AI actually is and what it can and cannot do. Crucially, generative AI is a seductive distraction from the type of AI that is most likely to make your life better, or even save it: Predictive AI. In contrast to AI designed for generative tasks, predictive AI involves tasks with a finite, known set of answers; the system just has to process information to say which answer is right. A basic example is plant recognition: Point your phone camera at a plant and learn that it’s a Western sword fern. Generative tasks, in contrast, have no finite set of correct answers: The system must blend snippets of information it’s been trained on to create, for example, a novel picture of a fern. 

The generative AI technology involved in chatbots, face-swaps, and synthetic video makes for stunning demos, driving clicks and sales as viewers run wild with ideas that superhuman AI will be capable of bringing us abundance or extinction. Yet predictive AI has quietly been improving weather prediction and food safety, enabling higher-quality music production, helping to organize photos, and accurately predicting the fastest driving routes. We incorporate predictive AI into our everyday lives without evening thinking about it, a testament to its indispensable utility.

To get a sense of the immense progress on predictive AI and its future potential, we can look at the trajectory of the past 20 years. In 2005, we couldn’t get AI to tell the difference between a person and a pencil. By 2013, AI still couldn’t reliably detect a bird in a photo, and the difference between a pedestrian and a Coke bottle was massively confounding (this is how I learned that bottles do kind of look like people, if people had no heads). The thought of deploying these systems in the real world was the stuff of science fiction. 

Yet over the past 10 years, predictive AI has not only nailed bird detection down to the specific species; it has rapidly improved life-critical medical services like identifying problematic lesions and heart arrhythmia. Because of this technology, seismologists can predict earthquakes and meteorologists can predict flooding more reliably than ever before. Accuracy has skyrocketed for consumer-facing tech that detects and classifies everything from what song you’re thinking of when you hum a tune to which objects to avoid while you’re driving—making self-driving cars a reality. 

In the very near future, we should be able to accurately detect tumors and forecast hurricanes long before they can hurt anyone, realizing the lifelong hopes of people all over the world. That might not be as flashy as generating your own Studio Ghibli–ish film, but it’s definitely hype-worthy. 

Predictive AI systems have also been shown to be incredibly useful when they leverage certain generative techniques within a constrained set of options. Systems of this type are diverse, spanning everything from outfit visualization to cross-language translation. Soon, predictive-generative hybrid systems will make it possible to clone your own voice speaking another language in real time, an extraordinary aid for travel (with serious impersonation risks). There’s considerable room for growth here, but generative AI delivers real value when anchored by strong predictive methods.

To understand the difference between these two broad classes of AI, imagine yourself as an AI system tasked with showing someone what a cat looks like. You could adopt a generative approach, cutting and pasting small fragments from various cat images (potentially from sources that object) to construct a seemingly perfect depiction. The ability of modern generative AI to produce such a flawless collage is what makes it so astonishing.

Alternatively, you could take the predictive approach: Simply locate and point to an existing picture of a cat. That method is much less glamorous but more energy-efficient and more likely to be accurate, and it properly acknowledges the original source. Generative AI is designed to create things that look real; predictive AI identifies what is real. A misunderstanding that generative systems are retrieving things when they are actually creating them has led to grave consequences when text is involved, requiring the withdrawal of legal rulings and the retraction of scientific articles.

Driving this confusion is a tendency for people to hype AI without making it clear what kind of AI they’re talking about (I reckon many don’t know). It’s very easy to equate “AI” with generative AI, or even just language-generating AI, and assume that all other capabilities fall out from there. That fallacy makes a ton of sense: The term literally references “intelligence,” and our human understanding of what “intelligence” might be is often mediated by the use of language. (Spoiler: No one actually knows what intelligence is.) But the phrase “artificial intelligence” was intentionally designed in the 1950s to inspire awe and allude to something humanlike. Today, it just refers to a set of disparate technologies for processing digital data. Some of my friends find it helpful to call it “mathy maths” instead.

The bias toward treating generative AI as the most powerful and real form of AI is troubling given that it consumes considerably more energy than predictive AI systems. It also means using existing human work in AI products against the original creators’ wishes and replacing human jobs with AI systems whose capabilities their work made possible in the first place—without compensation. AI can be amazingly powerful, but that doesn’t mean creators should be ripped off. 

Watching this unfold as an AI developer within the tech industry, I’ve drawn important lessons for next steps. The widespread appeal of AI is clearly linked to the intuitive nature of conversation-based interactions. But this method of engagement currently overuses generative methods where predictive ones would suffice, resulting in an awkward situation that’s confusing for users while imposing heavy costs in energy consumption, exploitation, and job displacement. 

We have witnessed just a glimpse of AI’s full potential: The current excitement around AI reflects what it could be, not what it is. Generation-based approaches strain resources while still falling short on representation, accuracy, and the wishes of people whose work is folded into the system. 

If we can shift the spotlight from the hype around generative technologies to the predictive advances already transforming daily life, we can build AI that is genuinely useful, equitable, and sustainable. The systems that help doctors catch diseases earlier, help scientists forecast disasters sooner, and help everyday people navigate their lives more safely are the ones poised to deliver the greatest impact. 

The future of beneficial AI will not be defined by the flashiest demos but by the quiet, rigorous progress that makes technology trustworthy. And if we build on that foundation—pairing predictive strength with more mature data practices and intuitive natural-language interfaces—AI can finally start living up to the promise that many people perceive today.

Dr. Margaret Mitchell is a computer science researcher and chief ethics scientist at AI startup Hugging Face. She has worked in the technology industry for 15 years, and has published over 100 papers on natural language generation, assistive technology, computer vision, and AI ethics. Her work has received numerous awards and has been implemented by multiple technology companies.

When the generative AI boom took off in 2022, Rudi Miller and her law school classmates were suddenly gripped with anxiety. “Before graduating, there was discussion about what the job market would look like for us if AI became adopted,” she recalls. 

So when it came time to choose a speciality, Miller—now a junior associate at the law firm Orrick—decided to become a litigator, the kind of lawyer who represents clients in court. She hoped the courtroom would be the last human stage. “Judges haven’t allowed ChatGPT-enabled robots to argue in court yet,” she says.

---

This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.

---

She had reason to be worried. The artificial-intelligence job apocalypse seemed to be coming for lawyers. In March 2023, researchers reported that GPT-4 had smashed the Uniform Bar Exam. That same month, an industry report predicted that 44% of legal work could be automated. The legal tech industry entered a boom as law firms began adopting generative AI to mine mountains of documents and draft contracts, work ordinarily done by junior associates. Last month, the law firm Clifford Chance axed 10% of its staff in London, citing increased use of AI as a reason.

But for all the hype, LLMs are still far from thinking like lawyers—let alone replacing them. The models continue to hallucinate case citations, struggle to navigate gray areas of the law and reason about novel questions, and stumble when they attempt to synthesize information scattered across statutes, regulations, and court cases. And there are deeper institutional reasons to think the models could struggle to supplant legal jobs. While AI is reshaping the grunt work of the profession, the end of lawyers may not be arriving anytime soon.

The big experiment

The legal industry has long been defined by long hours and grueling workloads, so the promise of superhuman efficiency is appealing. Law firms are experimenting with general-purpose tools like ChatGPT and Microsoft Copilot and specialized legal tools like Harvey and Thomson Reuters’ CoCounsel, with some building their own in-house tools on top of frontier models. They’re rolling out AI boot camps and letting associates bill hundreds of hours to AI experimentation. As of 2024, 47.8% of attorneys at law firms employing 500 or more lawyers used AI, according to the American Bar Association. 

But lawyers say that LLMs are a long way from reasoning well enough to replace them. Lucas Hale, a junior associate at McDermott Will & Schulte, has been embracing AI for many routine chores. He uses Relativity to sift through long documents and Microsoft Copilot for drafting legal citations. But when he turns to ChatGPT with a complex legal question, he finds the chatbot spewing hallucinations, rambling off topic, or drawing a blank.

“In the case where we have a very narrow question or a question of first impression for the court,” he says, referring to a novel legal question that a court has never decided before, “that’s the kind of thinking that the tool can’t do.”

Much of Lucas’s work involves creatively applying the law to new fact patterns. “Right now, I don’t think very much of the work that litigators do, at least not the work that I do, can be outsourced to an AI utility,” he says.

Allison Douglis, a senior associate at Jenner & Block, uses an LLM to kick off her legal research. But the tools only take her so far. “When it comes to actually fleshing out and developing an argument as a litigator, I don’t think they’re there,” she says. She has watched the models hallucinate case citations and fumble through ambiguous areas of the law.

“Right now, I would much rather work with a junior associate than an AI tool,” she says. “Unless they get extraordinarily good very quickly, I can’t imagine that changing in the near future.”

Beyond the bar

The legal industry has seemed ripe for an AI takeover ever since ChatGPT’s triumph on the bar exam. But passing a standardized test isn’t the same as practicing law. The exam tests whether people can memorize legal rules and apply them to hypothetical situations—not whether they can exercise strategic judgment in complicated realities or craft arguments in uncharted legal territory. And models can be trained to ace benchmarks without genuinely improving their reasoning.

But new benchmarks are aiming to better measure the models’ ability to do legal work in the real world. The Professional Reasoning Benchmark, published by ScaleAI in November, evaluated leading LLMs on legal and financial tasks designed by professionals in the field. The study found that the models have critical gaps in their reliability for professional adoption, with the best-performing model scoring only 37% on the most difficult legal problems, meaning it met just over a third of possible points on the evaluation criteria. The models frequently made inaccurate legal judgments, and if they did reach correct conclusions, they did so through incomplete or opaque reasoning processes. 

“The tools actually are not there to basically substitute [for] your lawyer,” says Afra Feyza Akyurek, the lead author of the paper. “Even though a lot of people think that LLMs have a good grasp of the law, it’s still lagging behind.” 

The paper builds on other benchmarks measuring the models’ performance on economically valuable work. The AI Productivity Index, published by the data firm Mercor in September and updated in December, found that the models have “substantial limitations” in performing legal work. The best-performing model scored 77.9% on legal tasks, meaning it satisfied roughly four out of five evaluation criteria. A model with such a score might generate substantial economic value in some industries, but in fields where errors are costly, it may not be useful at all, the early version of the study noted.  

Professional benchmarks are a big step forward in evaluating the LLMs’ real-world capabilities, but they may still not capture what lawyers actually do. “These questions, although more challenging than those in past benchmarks, still don’t fully reflect the kinds of subjective, extremely challenging questions lawyers tackle in real life,” says Jon Choi, a law professor at the University of Washington School of Law, who coauthored a study on legal benchmarks in 2023. 

Unlike math or coding, in which LLMs have made significant progress, legal reasoning may be challenging for the models to learn. The law deals with messy real-world problems, riddled with ambiguity and subjectivity, that often have no right answer, says Choi. Making matters worse, a lot of legal work isn’t recorded in ways that can be used to train the models, he says. When it is, documents can span hundreds of pages, scattered across statutes, regulations, and court cases that exist in a complex hierarchy.  

But a more fundamental limitation might be that LLMs are simply not trained to think like lawyers. “The reasoning models still don’t fully reason about problems like we humans do,” says Julian Nyarko, a law professor at Stanford Law School. The models may lack a mental model of the world—the ability to simulate a scenario and predict what will happen—and that capability could be at the heart of complex legal reasoning, he says. It’s possible that the current paradigm of LLMs trained on next-word prediction gets us only so far.  

The jobs remain

Despite early signs that AI is beginning to affect entry-level workers, labor statistics have yet to show that lawyers are being displaced. 93.4% of law school graduates in 2024 were employed within 10 months of graduation—the highest rate on record—according to the National Association for Law Placement. The number of graduates working in law firms rose by 13% from 2023 to 2024. 

For now, law firms are slow to shrink their ranks. “We’re not reducing headcounts at this point,” said Amy Ross, the chief of attorney talent at the law firm Ropes & Gray. 

Even looking ahead, the effects could be incremental. “I will expect some impact on the legal profession’s labor market, but not major,” says Mert Demirer, an economist at MIT. “AI is going to be very useful in terms of information discovery and summary,” he says, but for complex legal tasks, “the law’s low risk tolerance, plus the current capabilities of AI, are going to make that case less automatable at this point.” Capabilities may evolve over time, but that’s a big unknown.

It’s not just that the models themselves are not ready to replace junior lawyers. Institutional barriers may also shape how AI is deployed. Higher productivity reduces billable hours, challenging the dominant business model of law firms. Liability looms large for lawyers, and clients may still want a human on the hook. Regulations could also constrain how lawyers use the technology.

Still, as AI takes on some associate work, law firms may need to reinvent their training system. “When junior work dries up, you have to have a more formal way of teaching than hoping that an apprenticeship works,” says Ethan Mollick, a management professor at the Wharton School of the University of Pennsylvania.

Zach Couger, a junior associate at McDermott Will & Schulte, leans on ChatGPT to comb through piles of contracts he once slogged through by hand. He can’t imagine going back to doing the job himself, but he wonders what he’s missing. 

“I’m worried that I’m not getting the same reps that senior attorneys got,” he says, referring to the repetitive training that has long defined the early experiences of lawyers. “On the other hand, it is very nice to have a semi–knowledge expert to just ask questions to that’s not a partner who’s also very busy.” 

Even though an AI job apocalypse looks distant, the uncertainty sticks with him. Lately, Couger finds himself staying up late, wondering if he could be part of the last class of associates at big law firms: “I may be the last plane out.”

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.

In July, a widely cited MIT study claimed that 95% of organizations that invested in generative AI were getting “zero return.” Tech stocks briefly plunged. While the study itself was more nuanced than the headlines, for many it still felt like the first hard data point confirming what skeptics had muttered for months: Hype around AI might be outpacing reality.

Then, in August, OpenAI CEO Sam Altman said what everyone in Silicon Valley had been whispering. “Are we in a phase where investors as a whole are overexcited about AI?” he said during a press dinner I attended. “My opinion is yes.” 

---

This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.

---

He compared the current moment to the dot-com bubble. “When bubbles happen, smart people get overexcited about a kernel of truth,” he explained. “Tech was really important. The internet was a really big deal. People got overexcited.” 

With those comments, it was off to the races. The next day’s stock market dip was attributed to the sentiment he shared. The question “Are we in an AI bubble?” became inescapable.

Who thinks it is a bubble? 

The short answer: Lots of people. But not everyone agrees on who or what is overinflated. Tech leaders are using this moment of fear to take shots at their rivals and position themselves as clear winners on the other side. How they describe the bubble depends on where their company sits.

When I asked Meta CEO Mark Zuckerberg about the AI bubble in September, he ran through the historical analogies of past bubbles—railroads, fiber for the internet, the dot-com boom—and noted that in each case, “the infrastructure gets built out, people take on too much debt, and then you hit some blip … and then a lot of the companies end up going out of business.”

But Zuckerberg’s prescription wasn’t for Meta to pump the brakes. It was to keep spending: “If we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate, obviously. But I’d say the risk is higher on the other side.”

Bret Taylor, the chairman of OpenAI and CEO of the AI startup Sierra, uses a mental model from the late ’90s to help navigate this AI bubble. “I think the closest analogue to this AI wave is the dot-com boom or bubble, depending on your level of pessimism,” he recently told me. Back then, he explained, everyone knew e-commerce was going to be big, but there was a massive difference between Buy.com and Amazon. Taylor and others have been trying to position themselves as today’s Amazon.

Still others are arguing that the pain will be widespread. Google CEO Sundar Pichai told the BBC this month that there’s “some irrationality” in the current boom. Asked whether Google would be immune to a bubble bursting, he warned, “I think no company is going to be immune, including us.”

What’s inflating the bubble?

Companies are raising enormous sums of money and seeing unprecedented valuations. Much of that money, in turn, is going toward the buildout of massive data centers—on which both private companies like OpenAI and Elon Musk’s xAI and public ones such as Meta and Google are spending heavily. OpenAI has pledged that it will spend $500 billion to build AI data centers, more than 15 times what was spent on the Manhattan Project.

This eye-popping spending on AI data centers isn’t entirely detached from reality. The leaders of the top AI companies all stress that they’re bottlenecked by their limited access to computing power. You hear it constantly when you talk to them. Startups can’t get the GPU allocations they need. Hyperscalers are rationing compute, saving it for their best customers.

If today’s AI market is as brutally supply-constrained as tech leaders claim, perhaps aggressive infrastructure buildouts are warranted. But some of the numbers are too large to comprehend. Sam Altman has told employees that OpenAI’s moonshot goal is to build 250 gigawatts of computing capacity by 2033, roughly equaling India’s total national electricity demand. Such a plan would cost more than $12 trillion by today’s standards.

“I do think there’s real execution risk,” OpenAI president and cofounder Greg Brockman recently told me about the company’s aggressive infrastructure goals. “Everything we say about the future, we see that it’s a possibility. It is not a certainty, but I don’t think the uncertainty comes from scientific questions. It’s a lot of hard work.”

Who is exposed, and who is to blame?

It depends on who you ask. During the August press dinner, where he made his market-moving comments, Altman was blunt about where he sees the excess. He said it’s “insane” that some AI startups with “three people and an idea” are receiving funding at such high valuations. “That’s not rational behavior,” he said. “Someone’s gonna get burned there, I think.” As Safe Superintelligence cofounder (and former OpenAI chief scientist and cofounder) Ilya Sutskever put it on a recent podcast: Silicon Valley has “more companies than ideas.”

Demis Hassabis, the CEO of Google DeepMind, offered a similar diagnosis when I spoke with him in November. “It feels like there’s obviously a bubble in the private market,” he said. “You look at seed rounds with just nothing being tens of billions of dollars. That seems a little unsustainable.”

Anthropic CEO Dario Amodei also struck at his competition during the New York Times DealBook Summit in early December. He said he feels confident about the technology itself but worries about how others are behaving on the business side: “On the economic side, I have my concerns where, even if the technology fulfills all its promises, I think there are players in the ecosystem who, if they just make a timing error, they just get it off by a little bit, bad things could happen.”

He stopped short of naming Sam Altman and OpenAI, but the implication was clear. “There are some players who are YOLOing,” he said. “Let’s say you’re a person who just kind of constitutionally wants to YOLO things or just likes big numbers. Then you may turn the dial too far.”

Amodei also flagged “circular deals,” or the increasingly common arrangements where chip suppliers like Nvidia invest in AI companies that then turn around and spend those funds on their chips. Anthropic has done some of these, he said, though “not at the same scale as some other players.” (OpenAI is at the center of a number of such deals, as are Nvidia, CoreWeave, and a roster of other players.) 

The danger, he explained, comes when the numbers get too big: “If you start stacking these where they get to huge amounts of money, and you’re saying, ’By 2027 or 2028 I need to make $200 billion a year,’ then yeah, you can overextend yourself.”

Zuckerberg shared a similar message at an internal employee Q&A session after Meta’s last earnings call. He noted that unprofitable startups like OpenAI and Anthropic risk bankruptcy if they misjudge the timing of their investments, but Meta has the advantage of strong cash flow, he reassured staff.

How could a bubble burst?

My conversations with tech executives and investors suggest that the bubble will be most likely to pop if overfunded startups can’t turn a profit or grow into their lofty valuations. This bubble could last longer than than past ones, given that private markets aren’t traded on public markets and therefore move more slowly, but the ripple effects will still be profound when the end comes. 

If companies making grand commitments to data center buildouts no longer have the revenue growth to support them, the headline deals that have propped up the stock market come into question. Anthropic’s Amodei illustrated the problem during his DealBook Summit appearance, where he said the multi-year data center commitments he has to make combine with the company’s rapid, unpredictable revenue growth rate to create a “cone of uncertainty” about how much to spend.

The two most prominent private players in AI, OpenAI and Anthropic, have yet to turn a profit. A recent Deutsche Bank chart put the situation in stark historical context. Amazon burned through $3 billion before becoming profitable. Tesla, around $4 billion. Uber, $30 billion. OpenAI is projected to burn through $140 billion by 2029, while Anthropic is expected to burn $20 billion by 2027.

Consultants at Bain estimate that the wave of AI infrastructure spending will require $2 trillion in annual AI revenue by 2030 just to justify the investment. That’s more than the combined 2024 revenue of Amazon, Apple, Alphabet, Microsoft, Meta, and Nvidia. When I talk to leaders of these large tech companies, they all agree that their sprawling businesses can absorb an expensive miscalculation about the returns from their AI infrastructure buildouts. It’s all the other companies that are either highly leveraged with debt or just unprofitable—even OpenAI and Anthropic—that they worry about. 

Still, given the level of spending on AI, it still needs a viable business model beyond subscriptions, which won’t be able to  drive profits from billions of people’s eyeballs like the ad-driven businesses that have defined the last 20 years of the internet. Even the largest tech companies know they need to ship the world-changing agents they keep hyping: AI that can fully replace coworkers and complete tasks in the real world.

For now, investors are mostly buying into the hype of the powerful AI systems that these data center buildouts will supposedly unlock in the future. At some point the biggest spenders, like OpenAI, will need to show investors that the money spent on the infrastructure buildout was worth it.

There’s also still a lot of uncertainty about the technical direction that AI is heading in. LLMs are expected to remain critical to more advanced AI systems, but industry leaders can’t seem to agree on which additional breakthroughs are needed to achieve artificial general intelligence, or AGI. Some are betting on new kinds of AI that can understand the physical world, while others are focused on training AI to learn in a general way, like a human. In other words, what if all this unprecedented spending turns out to have been backing the wrong horse?

The question now

What makes this moment surreal is the honesty. The same people pouring billions into AI will openly tell you it might all come crashing down. 

Taylor framed it as two truths existing at once. “I think it is both true that AI will transform the economy,” he told me, “and I think we’re also in a bubble, and a lot of people will lose a lot of money. I think both are absolutely true at the same time.”

He compared it to the internet. Webvan failed, but Instacart succeeded years later with essentially the same idea. If you were an Amazon shareholder from its IPO to now, you’re looking pretty good. If you were a Webvan shareholder, you probably feel differently. 

“When the dust settles and you see who the winners are, society benefits from those inventions,” Amazon founder Jeff Bezos said in October. “This is real. The benefit to society from AI is going to be gigantic.”

Goldman Sachs says the AI boom now looks the way tech stocks did in 1997, several years before the dot-com bubble actually burst. The bank flagged five warning signs seen in the late 1990s that investors should watch now: peak investment spending, falling corporate profits, rising corporate debt, Fed rate cuts, and widening credit spreads. We’re probably not at 1999 levels yet. But the imbalances are building fast. Michael Burry, who famously called the 2008 housing bubble collapse (as seen in the film The Big Short), recently compared the AI boom to the 1990s dot-com bubble too.

Maybe AI will save us from our own irrational exuberance. But for now, we’re living in an in-between moment when everyone knows what’s coming but keeps blowing more air into the balloon anyway. As Altman put it that night at dinner: “Someone is going to lose a phenomenal amount of money. We don’t know who.”

Alex Heath is the author of Sources, a newsletter about the AI race, and the cohost of ACCESS, a podcast about the tech industry’s inside conversations. Previously, he was deputy editor at The Verge.

The microwave-size instrument at Lila Sciences in Cambridge, Massachusetts, doesn’t look all that different from others that I’ve seen in state-of-the-art materials labs. Inside its vacuum chamber, the machine zaps a palette of different elements to create vaporized particles, which then fly through the chamber and land to create a thin film, using a technique called sputtering. What sets this instrument apart is that artificial intelligence is running the experiment; an AI agent, trained on vast amounts of scientific literature and data, has determined the recipe and is varying the combination of elements. 

Later, a person will walk the samples, each containing multiple potential catalysts, over to a different part of the lab for testing. Another AI agent will scan and interpret the data, using it to suggest another round of experiments to try to optimize the materials’ performance.  

---

This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.

---

For now, a human scientist keeps a close eye on the experiments and will approve the next steps on the basis of the AI’s suggestions and the test results. But the startup is convinced this AI-controlled machine is a peek into the future of materials discovery—one in which autonomous labs could make it far cheaper and faster to come up with novel and useful compounds. 

Flush with hundreds of millions of dollars in new funding, Lila Sciences is one of AI’s latest unicorns. The company is on a larger mission to use AI-run autonomous labs for scientific discovery—the goal is to achieve what it calls scientific superintelligence. But I’m here this morning to learn specifically about the discovery of new materials. 

We desperately need better materials to solve our problems. We’ll need improved electrodes and other parts for more powerful batteries; compounds to more cheaply suck carbon dioxide out of the air; and better catalysts to make green hydrogen and other clean fuels and chemicals. And we will likely need novel materials like higher-temperature superconductors, improved magnets, and different types of semiconductors for a next generation of breakthroughs in everything from quantum computing to fusion power to AI hardware. 

But materials science has not had many commercial wins in the last few decades. In part because of its complexity and the lack of successes, the field has become something of an innovation backwater, overshadowed by the more glamorous—and lucrative—search for new drugs and insights into biology.

The idea of using AI for materials discovery is not exactly new, but it got a huge boost in 2020 when DeepMind showed that its AlphaFold2 model could accurately predict the three-dimensional structure of proteins. Then, in 2022, came the success and popularity of ChatGPT. The hope that similar AI models using deep learning could aid in doing science captivated tech insiders. Why not use our new generative AI capabilities to search the vast chemical landscape and help simulate atomic structures, pointing the way to new substances with amazing properties?

“Simulations can be super powerful for framing problems and understanding what is worth testing in the lab. But there’s zero problems we can ever solve in the real world with simulation alone.”

John Gregoire, Lila Sciences, chief autonomous science officer

Researchers touted an AI model that had reportedly discovered “millions of new materials.” The money began pouring in, funding a host of startups. But so far there has been no “eureka” moment, no ChatGPT-like breakthrough—no discovery of new miracle materials or even slightly better ones.

The startups that want to find useful new compounds face a common bottleneck: By far the most time-consuming and expensive step in materials discovery is not imagining new structures but making them in the real world. Before trying to synthesize a material, you don’t know if, in fact, it can be made and is stable, and many of its properties remain unknown until you test it in the lab.

“Simulations can be super powerful for kind of framing problems and understanding what is worth testing in the lab,” says John Gregoire, Lila Sciences’ chief autonomous science officer. “But there’s zero problems we can ever solve in the real world with simulation alone.” 

Startups like Lila Sciences have staked their strategies on using AI to transform experimentation and are building labs that use agents to plan, run, and interpret the results of experiments to synthesize new materials. Automation in laboratories already exists. But the idea is to have AI agents take it to the next level by directing autonomous labs, where their tasks could include designing experiments and controlling the robotics used to shuffle samples around. And, most important, companies want to use AI to vacuum up and analyze the vast amount of data produced by such experiments in the search for clues to better materials.

If they succeed, these companies could shorten the discovery process from decades to a few years or less, helping uncover new materials and optimize existing ones. But it’s a gamble. Even though AI is already taking over many laboratory chores and tasks, finding new—and useful—materials on its own is another matter entirely. 

Innovation backwater

I have been reporting about materials discovery for nearly 40 years, and to be honest, there have been only a few memorable commercial breakthroughs, such as lithium-­ion batteries, over that time. There have been plenty of scientific advances to write about, from perovskite solar cells to graphene transistors to metal-­organic frameworks (MOFs), materials based on an intriguing type of molecular architecture that recently won its inventors a Nobel Prize. But few of those advances—including MOFs—have made it far out of the lab. Others, like quantum dots, have found some commercial uses, but in general, the kinds of life-changing inventions created in earlier decades have been lacking. 

Blame the amount of time (typically 20 years or more) and the hundreds of millions of dollars it takes to make, test, optimize, and manufacture a new material—and the industry’s lack of interest in spending that kind of time and money in low-margin commodity markets. Or maybe we’ve just run out of ideas for making stuff.

The need to both speed up that process and find new ideas is the reason researchers have turned to AI. For decades, scientists have used computers to design potential materials, calculating where to place atoms to form structures that are stable and have predictable characteristics. It’s worked—but only kind of. Advances in AI have made that computational modeling far faster and have promised the ability to quickly explore a vast number of possible structures. Google DeepMind, Meta, and Microsoft have all launched efforts to bring AI tools to the problem of designing new materials. 

But the limitations that have always plagued computational modeling of new materials remain. With many types of materials, such as crystals, useful characteristics often can’t be predicted solely by calculating atomic structures.

To uncover and optimize those properties, you need to make something real. Or as Rafael Gómez-Bombarelli, one of Lila’s cofounders and an MIT professor of materials science, puts it: “Structure helps us think about the problem, but it’s neither necessary nor sufficient for real materials problems.”

Perhaps no advance exemplified the gap between the virtual and physical worlds more than DeepMind’s announcement in late 2023 that it had used deep learning to discover “millions of new materials,” including 380,000 crystals that it declared “the most stable, making them promising candidates for experimental synthesis.” In technical terms, the arrangement of atoms represented a minimum energy state where they were content to stay put. This was “an order-of-magnitude expansion in stable materials known to humanity,” the DeepMind researchers proclaimed.

To the AI community, it appeared to be the breakthrough everyone had been waiting for. The DeepMind research not only offered a gold mine of possible new materials, it also created powerful new computational methods for predicting a large number of structures.

But some materials scientists had a far different reaction. After closer scrutiny, researchers at the University of California, Santa Barbara, said they’d found “scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility.” In fact, the scientists reported, they didn’t find any truly novel compounds among the ones they looked at; some were merely “trivial” variations of known ones. The scientists appeared particularly peeved that the potential compounds were labeled materials. They wrote: “We would respectfully suggest that the work does not report any new materials but reports a list of proposed compounds. In our view, a compound can be called a material when it exhibits some functionality and, therefore, has potential utility.”

Some of the imagined crystals simply defied the conditions of the real world. To do computations on so many possible structures, DeepMind researchers simulated them at absolute zero, where atoms are well ordered; they vibrate a bit but don’t move around. At higher temperatures—the kind that would exist in the lab or anywhere in the world—the atoms fly about in complex ways, often creating more disorderly crystal structures. A number of the so-called novel materials predicted by DeepMind appeared to be well-ordered versions of disordered ones that were already known. 

More generally, the DeepMind paper was simply another reminder of how challenging it is to capture physical realities in virtual simulations—at least for now. Because of the limitations of computational power, researchers typically perform calculations on relatively few atoms. Yet many desirable properties are determined by the microstructure of the materials—at a scale much larger than the atomic world. And some effects, like high-temperature superconductivity or even the catalysis that is key to many common industrial processes, are far too complex or poorly understood to be explained by atomic simulations alone.

A common language

Even so, there are signs that the divide between simulations and experimental work is beginning to narrow. DeepMind, for one, says that since the release of the 2023 paper it has been working with scientists in labs around the world to synthesize AI-identified compounds and has achieved some success. Meanwhile, a number of the startups entering the space are looking to combine computational and experimental expertise in one organization. 

One such startup is Periodic Labs, cofounded by Ekin Dogus Cubuk, a physicist who led the scientific team that generated the 2023 DeepMind headlines, and by Liam Fedus, a co-creator of ChatGPT at OpenAI. Despite its founders’ background in computational modeling and AI software, the company is building much of its materials discovery strategy around synthesis done in automated labs. 

The vision behind the startup is to link these different fields of expertise by using large language models that are trained on scientific literature and able to learn from ongoing experiments. An LLM might suggest the recipe and conditions to make a compound; it can also interpret test data and feed additional suggestions to the startup’s chemists and physicists. In this strategy, simulations might suggest possible material candidates, but they are also used to help explain the experimental results and suggest possible structural tweaks.

The grand prize would be a room-temperature superconductor, a material that could transform computing and electricity but that has eluded scientists for decades.

Periodic Labs, like Lila Sciences, has ambitions beyond designing and making new materials. It wants to “create an AI scientist”—specifically, one adept at the physical sciences. “LLMs have gotten quite good at distilling chemistry information, physics information,” says Cubuk, “and now we’re trying to make it more advanced by teaching it how to do science—for example, doing simulations, doing experiments, doing theoretical modeling.”

The approach, like that of Lila Sciences, is based on the expectation that a better understanding of the science behind materials and their synthesis will lead to clues that could help researchers find a broad range of new ones. One target for Periodic Labs is materials whose properties are defined by quantum effects, such as new types of magnets. The grand prize would be a room-temperature superconductor, a material that could transform computing and electricity but that has eluded scientists for decades.

Superconductors are materials in which electricity flows without any resistance and, thus, without producing heat. So far, the best of these materials become superconducting only at relatively low temperatures and require significant cooling. If they can be made to work at or close to room temperature, they could lead to far more efficient power grids, new types of quantum computers, and even more practical high-speed magnetic-levitation trains. 

The failure to find a room-­temperature superconductor is one of the great disappointments in materials science over the last few decades. I was there when President Reagan spoke about the technology in 1987, during the peak hype over newly made ceramics that became superconducting at the relatively balmy temperature of 93 Kelvin (that’s −292 °F), enthusing that they “bring us to the threshold of a new age.” There was a sense of optimism among the scientists and businesspeople in that packed ballroom at the Washington Hilton as Reagan anticipated “a host of benefits, not least among them a reduced dependence on foreign oil, a cleaner environment, and a stronger national economy.” In retrospect, it might have been one of the last times that we pinned our economic and technical aspirations on a breakthrough in materials.

The promised new age never came. Scientists still have not found a material that becomes superconducting at room temperatures, or anywhere close, under normal conditions. The best existing superconductors are brittle and tend to make lousy wires.

One of the reasons that finding higher-­temperature superconductors has been so difficult is that no theory explains the effect at relatively high temperatures—or can predict it simply from the placement of atoms in the structure. It will ultimately fall to lab scientists to synthesize any interesting candidates, test them, and search the resulting data for clues to understanding the still puzzling phenomenon. Doing so, says Cubuk, is one of the top priorities of Periodic Labs. 

AI in charge

It can take a researcher a year or more to make a crystal structure for the first time. Then there are typically years of further work to test its properties and figure out how to make the larger quantities needed for a commercial product. 

Startups like Lila Sciences and Periodic Labs are pinning their hopes largely on the prospect that AI-directed experiments can slash those times. One reason for the optimism is that many labs have already incorporated a lot of automation, for everything from preparing samples to shuttling test items around. Researchers routinely use robotic arms, software, automated versions of microscopes and other analytical instruments, and mechanized tools for manipulating lab equipment.

The automation allows, among other things, for high-throughput synthesis, in which multiple samples with various combinations of ingredients are rapidly created and screened in large batches, greatly speeding up the experiments.

The idea is that using AI to plan and run such automated synthesis can make it far more systematic and efficient. AI agents, which can collect and analyze far more data than any human possibly could, can use real-time information to vary the ingredients and synthesis conditions until they get a sample with the optimal properties. Such AI-directed labs could do far more experiments than a person and could be far smarter than existing systems for high-throughput synthesis. 

But so-called self-driving labs for materials are still a work in progress.

Many types of materials require solid-­state synthesis, a set of processes that are far more difficult to automate than the liquid-­handling activities that are commonplace in making drugs. You need to prepare and mix powders of multiple inorganic ingredients in the right combination for making, say, a catalyst and then decide how to process the sample to create the desired structure—for example, identifying the right temperature and pressure at which to carry out the synthesis. Even determining what you’ve made can be tricky.

In 2023, the A-Lab at Lawrence Berkeley National Laboratory claimed to be the first fully automated lab to use inorganic powders as starting ingredients. Subsequently, scientists reported that the autonomous lab had used robotics and AI to synthesize and test 41 novel materials, including some predicted in the DeepMind database. Some critics questioned the novelty of what was produced and complained that the automated analysis of the materials was not up to experimental standards, but the Berkeley researchers defended the effort as simply a demonstration of the autonomous system’s potential.

“How it works today and how we envision it are still somewhat different. There’s just a lot of tool building that needs to be done,” says Gerbrand Ceder, the principal scientist behind the A-Lab. 

AI agents are already getting good at doing many laboratory chores, from preparing recipes to interpreting some kinds of test data—finding, for example, patterns in a micrograph that might be hidden to the human eye. But Ceder is hoping the technology could soon “capture human decision-making,” analyzing ongoing experiments to make strategic choices on what to do next. For example, his group is working on an improved synthesis agent that would better incorporate what he calls scientists’ “diffused” knowledge—the kind gained from extensive training and experience. “I imagine a world where people build agents around their expertise, and then there’s sort of an uber-model that puts it together,” he says. “The uber-model essentially needs to know what agents it can call on and what they know, or what their expertise is.”

“In one field that I work in, solid-state batteries, there are 50 papers published every day. And that is just one field that I work in. The A I revolution is about finally gathering all the scientific data we have.”

Gerbrand Ceder, principal scientist, A-Lab

One of the strengths of AI agents is their ability to devour vast amounts of scientific literature. “In one field that I work in, solid-­state batteries, there are 50 papers published every day. And that is just one field that I work in,” says Ceder. It’s impossible for anyone to keep up. “The AI revolution is about finally gathering all the scientific data we have,” he says. 

Last summer, Ceder became the chief science officer at an AI materials discovery startup called Radical AI and took a sabbatical from the University of California, Berkeley, to help set up its self-driving labs in New York City. A slide deck shows the portfolio of different AI agents and generative models meant to help realize Ceder’s vision. If you look closely, you can spot an LLM called the “orchestrator”—it’s what CEO Joseph Krause calls the “head honcho.” 

New hope

So far, despite the hype around the use of AI to discover new materials and the growing momentum—and money—behind the field, there still has not been a convincing big win. There is no example like the 2016 victory of DeepMind’s AlphaGo over a Go world champion. Or like AlphaFold’s achievement in mastering one of biomedicine’s hardest and most time-consuming chores, predicting 3D structures of proteins. 

The field of materials discovery is still waiting for its moment. It could come if AI agents can dramatically speed the design or synthesis of practical materials, similar to but better than what we have today. Or maybe the moment will be the discovery of a truly novel one, such as a room-­temperature superconductor.

With or without such a breakthrough moment, startups face the challenge of trying to turn their scientific achievements into useful materials. The task is particularly difficult because any new materials would likely have to be commercialized in an industry dominated by large incumbents that are not particularly prone to risk-taking.

Susan Schofer, a tech investor and partner at the venture capital firm SOSV, is cautiously optimistic about the field. But Schofer, who spent several years in the mid-2000s as a catalyst researcher at one of the first startups using automation and high-throughput screening for materials discovery (it didn’t survive), wants to see some evidence that the technology can translate into commercial successes when she evaluates startups to invest in.  

In particular, she wants to see evidence that the AI startups are already “finding something new, that’s different, and know how they are going to iterate from there.” And she wants to see a business model that captures the value of new materials. She says, “I think the ideal would be: I got a spec from the industry. I know what their problem is. We’ve defined it. Now we’re going to go build it. Now we have a new material that we can sell, that we have scaled up enough that we’ve proven it. And then we partner somehow to manufacture it, but we get revenue off selling the material.”

Schofer says that while she gets the vision of trying to redefine science, she’d advise startups to “show us how you’re going to get there.” She adds, “Let’s see the first steps.”

Demonstrating those first steps could be essential in enticing large existing materials companies to embrace AI technologies more fully. Corporate researchers in the industry have been burned before—by the promise over the decades that increasingly powerful computers will magically design new materials; by combinatorial chemistry, a fad that raced through materials R&D labs in the early 2000s with little tangible result; and by the promise that synthetic biology would make our next generation of chemicals and materials.

More recently, the materials community has been blanketed by a new hype cycle around AI. Some of that hype was fueled by the 2023 DeepMind announcement of the discovery of “millions of new materials,” a claim that, in retrospect, clearly overpromised. And it was further fueled when an MIT economics student posted a paper in late 2024 claiming that a large, unnamed corporate R&D lab had used AI to efficiently invent a slew of new materials. AI, it seemed, was already revolutionizing the industry.

A few months later, the MIT economics department concluded that “the paper should be withdrawn from public discourse.” Two prominent MIT economists who are acknowledged in a footnote in the paper added that they had “no confidence in the provenance, reliability or validity of the data and the veracity of the research.”

Can AI move beyond the hype and false hopes and truly transform materials discovery? Maybe. There is ample evidence that it’s changing how materials scientists work, providing them—if nothing else—with useful lab tools. Researchers are increasingly using LLMs to query the scientific literature and spot patterns in experimental data. 

But it’s still early days in turning those AI tools into actual materials discoveries. The use of AI to run autonomous labs, in particular, is just getting underway; making and testing stuff takes time and lots of money. The morning I visited Lila Sciences, its labs were largely empty, and it’s now preparing to move into a much larger space a few miles away. Periodic Labs is just beginning to set up its lab in San Francisco. It’s starting with manual synthesis guided by AI predictions; its robotic high-throughput lab will come soon. Radical AI reports that its lab is almost fully autonomous but plans to soon move to a larger space.

When I talk to the scientific founders of these startups, I hear a renewed excitement about a field that long operated in the shadows of drug discovery and genomic medicine. For one thing, there is the money. “You see this enormous enthusiasm to put AI and materials together,” says Ceder. “I’ve never seen this much money flow into materials.”

Reviving the materials industry is a challenge that goes beyond scientific advances, however. It means selling companies on a whole new way of doing R&D.

But the startups benefit from a huge dose of confidence borrowed from the rest of the AI industry. And maybe that, after years of playing it safe, is just what the materials business needs.