Soon after OpenAI released GPT-4o on Monday, May 13, some Chinese speakers started to notice that something seemed off about this newest version of the chatbot: the tokens it uses to parse text were full of spam and porn phrases.

On May 14, Tianle Cai, a PhD student at Princeton University studying inference efficiency in large language models like those that power such chatbots, accessed GPT-4o’s public token library and pulled a list of the 100 longest Chinese tokens the model uses to parse and compress Chinese prompts. 

Humans read in words, but LLMs read in tokens, which are distinct units in a sentence that have consistent and significant meanings. Besides dictionary words, they also include suffixes, common expressions, names, and more. The more tokens a model encodes, the faster the model can “read” a sentence and the less computing power it consumes, thus making the response cheaper.

Of the 100 results, only three of them are common enough to be used in everyday conversations; everything else consisted of words and expressions used specifically in the contexts of either gambling or pornography. The longest token, lasting 10.5 Chinese characters, literally means “_free Japanese porn video to watch.” Oops.

“This is sort of ridiculous,” Cai wrote, and he posted the list of tokens on GitHub.

OpenAI did not respond to questions sent by MIT Technology Review prior to publication.

GPT-4o is supposed to be better than its predecessors at handling multi-language tasks. In particular, the advances are achieved through a new tokenization tool that does a better job compressing texts in non-English languages.

But at least when it comes to the Chinese language, the new tokenizer used by GPT-4o has introduced a disproportionate number of meaningless phrases. Experts say that’s likely due to insufficient data cleaning and filtering before the tokenizer was trained. 

Because these tokens are not actual commonly spoken words or phrases, the chatbot can fail to grasp their meanings. Researchers have been able to leverage that and trick GPT-4o into hallucinating answers or even circumventing the safety guardrails OpenAI had put in place.

Why non-English tokens matter

The easiest way for a model to process text is character by character, but that’s obviously more time consuming and laborious than recognizing that a certain string of characters—like “c-r-y-p-t-o-c-u-r-r-e-n-c-y”—always means the same thing. These series of characters are encoded as “tokens” the model can use to process prompts. Including more and longer tokens usually means the LLMs are more efficient and affordable for users—who are often billed per token.

When OpenAI released GPT-4o on May 13, it also released a new tokenizer to replace the one it used in previous versions, GPT-3.5 and GPT-4. The new tokenizer especially adds support for non-English languages, according to OpenAI’s website.

The new tokenizer has 200,000 tokens in total, and about 25% are in non-English languages, says Deedy Das, an AI investor at Menlo Ventures. He used language filters to count the number of tokens in different languages, and the top languages, besides English, are Russian, Arabic, and Vietnamese.

“So the tokenizer’s main impact, in my opinion, is you get the cost down in these languages, not that the quality in these languages goes dramatically up,” Das says. When an LLM has better and longer tokens in non-English languages, it can analyze the prompts faster and charge users less for the same answer. With the new tokenizer, “you’re looking at almost four times cost reduction,” he says.

Das, who also speaks Hindi and Bengali, took a look at the longest tokens in those languages. The tokens reflect discussions happening in those languages, so they include words like “Narendra” or “Pakistan,” but common English terms like “Prime Minister,” “university,” and “international” also come up frequently. They also don’t exhibit the issues surrounding the Chinese tokens.

That likely reflects the training data in those languages, Das says: “My working theory is the websites in Hindi and Bengali are very rudimentary. It’s like [mostly] news articles. So I would expect this to be the case. There are not many spam bots and porn websites trying to happen in these languages. It’s mostly going to be in English.”

Polluted data and a lack of cleaning

However, things are drastically different in Chinese. According to multiple researchers who have looked into the new library of tokens used for GPT-4o, the longest tokens in Chinese are almost exclusively spam words used in pornography, gambling, and scamming contexts. Even shorter tokens, like three-character-long Chinese words, reflect those topics to a significant degree.

“The problem is clear: the corpus used to train [the tokenizer] is not clean. The English tokens seem fine, but the Chinese ones are not,” says Cai from Princeton University. It is not rare for a language model to crawl spam when collecting training data, but usually there will be significant effort taken to clean up the data before it’s used. “It’s possible that they didn’t do proper data clearing when it comes to Chinese,” he says.

The content of these Chinese tokens could suggest that they have been polluted by a specific phenomenon: websites hijacking unrelated content in Chinese or other languages to boost spam messages. 

These messages are often advertisements for pornography videos and gambling websites. They could be real businesses or merely scams. And the language is inserted into content farm websites or sometimes legitimate websites so they can be indexed by search engines, circumvent the spam filters, and come up in random searches. For example, Google indexed one search result page on a US National Institutes of Health website, which lists a porn site in Chinese. The same site name also appeared in at least five Chinese tokens in GPT-4o. 

Chinese users have reported that these spam sites appeared frequently in unrelated Google search results this year, including in comments made to Google Search’s support community. It’s likely that these websites also found their way into OpenAI’s training database for GPT-4o’s new tokenizer. 

The same issue didn’t exist with the previous-generation tokenizer and Chinese tokens used for GPT-3.5 and GPT-4, says Zhengyang Geng, a PhD student in computer science at Carnegie Mellon University. There, the longest Chinese tokens are common terms like “life cycles” or “auto-generation.” 

Das, who worked on the Google Search team for three years, says the prevalence of spam content is a known problem and isn’t that hard to fix. “Every spam problem has a solution. And you don’t need to cover everything in one technique,” he says. Even simple solutions like requesting an automatic translation of the content when detecting certain keywords could “get you 60% of the way there,” he adds.

But OpenAI likely didn’t clean the Chinese data set or the tokens before the release of GPT-4o, Das says:  “At the end of the day, I just don’t think they did the work in this case.”

It’s unclear whether any other languages are affected. One X user reported that a similar prevalence of porn and gambling content in Korean tokens.

The tokens can be used to jailbreak

Users have also found that these tokens can be used to break the LLM, either getting it to spew out completely unrelated answers or, in rare cases, to generate answers that are not allowed under OpenAI’s safety standards.

Geng of Carnegie Mellon University asked GPT-4o to translate some of the long Chinese tokens into English. The model then proceeded to translate words that were never included in the prompts, a typical result of LLM hallucinations.

He also succeeded in using the same tokens to “jailbreak” GPT-4o—that is, to get the model to generate things it shouldn’t. “It’s pretty easy to use these [rarely used] tokens to induce undefined behaviors from the models,” Geng says. “I did some personal red-teaming experiments … The simplest example is asking it to make a bomb. In a normal condition, it would decline it, but if you first use these rare words to jailbreak it, then it will start following your orders. Once it starts to follow your orders, you can ask it all kinds of questions.”

In his tests, which Geng chooses not to share with the public, he says he can see GPT-4o generating the answers line by line. But when it almost reaches the end, another safety mechanism kicks in, detects unsafe content, and blocks it from being shown to the user.

The phenomenon is not unusual in LLMs, says Sander Land, a machine-learning engineer at Cohere, a Canadian AI company. Land and his colleague Max Bartolo recently drafted a paper on how to detect the unusual tokens that can be used to cause models to glitch. One of the most famous examples was “_SolidGoldMagikarp,” a Reddit username that was found to get ChatGPT to generate unrelated, weird, and unsafe answers.

The problem lies in the fact that sometimes the tokenizer and the actual LLM are trained on different data sets, and what was prevalent in the tokenizer data set is not in the LLM data set for whatever reason. The result is that while the tokenizer picks up certain words that it sees frequently, the model is not sufficiently trained on them and never fully understands what these “under-trained” tokens mean. In the _SolidGoldMagikarp case, the username was likely included in the tokenizer training data but not in the actual GPT training data, leaving GPT at a loss about what to do with the token. “And if it has to say something … it gets kind of a random signal and can do really strange things,” Land says.

And different models could glitch differently in this situation. “Like, Llama 3 always gives back empty space but sometimes then talks about the empty space as if there was something there. With other models, I think Gemini, when you give it one of these tokens, it provides a beautiful essay about aluminum, and [the question] didn’t have anything to do with aluminum,” says Land.

To solve this problem, the data set used for training the tokenizer should well represent the data set for the LLM, he says, so there won’t be mismatches between them. If the actual model has gone through safety filters to clean out porn or spam content, the same filters should be applied to the tokenizer data. In reality, this is sometimes hard to do because training LLMs takes months and involves constant improvement, with spam content being filtered out, while token training is usually done at an early stage and may not involve the same level of filtering. 

While experts agree it’s not too difficult to solve the issue, it could get complicated as the result gets looped into multi-step intra-model processes, or when the polluted tokens and models get inherited in future iterations. For example, it’s not possible to publicly test GPT-4o’s video and audio functions yet, and it’s unclear whether they suffer from the same glitches that can be caused by these Chinese tokens.

“The robustness of visual input is worse than text input in multimodal models,” says Geng, whose research focus is on visual models. Filtering a text data set is relatively easy, but filtering visual elements will be even harder. “The same issue with these Chinese spam tokens could become bigger with visual tokens,” he says.

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

We finally know the result of a legal case I’ve been tracking in Hong Kong for almost a year. Last week, the Hong Kong Court of Appeal granted an injunction that permits the city government to go to Western platforms like YouTube and Spotify and demand they remove the protest anthem “Glory to Hong Kong,” because the government claims it has been used for sedition.

To read more about how this injunction is specifically designed for Western Big Tech platforms, and the impact it’s likely to have on internet freedom, you can read my story here.

Aside from the depressing implications for pro-democracy movements’ decline in Hong Kong, this lawsuit has also been an interesting case study of the local government’s complicated relationship with internet control and censorship.

I was following this case because it’s a perfect example of how censorship can be built brick by brick. Having reported on China for so long, I sometimes take for granted how powerful and all-encompassing its censorship regime is and need to be reminded that the same can’t be said for most other places in the world.

Hong Kong had a free internet in the past. And unlike mainland China, it remains relatively open: almost all Western platforms and services are still available there, and only a few websites have been censored in recent years. 

Since Hong Kong was returned to China from the UK in 1997, the Chinese central government has clashed several times with local pro-democracy movements asking for universal elections and less influence from Beijing. As a result, it started cementing tighter and tighter control over Hong Kong, and people have been worrying about whether its Great Firewall will eventually extend there. But actually, neither Beijing nor Hong Kong may want to see that happen. All the recent legal maneuverings are only necessary because the government doesn’t want a full-on ban of Western platforms.

When I visited Hong Kong last November, it was pretty clear that both Beijing and Hong Kong want to take advantage of the free flow of finance and business through the city. That’s why the Hong Kong government was given tacit permission in 2023 to explore government cryptocurrency projects, even though crypto trading and mining are illegal in China. Hong Kong officials have boasted on many occasions about the city’s value proposition: connecting untapped demand in the mainland to the wider crypto world by attracting mainland investors and crypto companies to set up shop in Hong Kong. 

But that wouldn’t be possible if Hong Kong closed off its internet. Imagine a “global” crypto industry that couldn’t access Twitter or Discord. Crypto is only one example, but the things that have made Hong Kong successful—the nonstop exchange of cargo, capital, ideas, and people—would cease to function if basic and universal tools like Google or Facebook became unavailable.

That’s why there are these calculated offenses on internet freedom in Hong Kong. It’s about seeking control but also leaving some breathing space; it’s as much about looking tough on the outside as negotiating with platforms down below; it’s about showing its determination to Beijing but also not showing too much aggression to the West. 

For example, the experts I’ve talked to don’t expect the government to request that YouTube remove the videos for everyone globally. More likely, they may ask for the content to be geo-blocked just for users in Hong Kong.

“As long as Hong Kong is still useful as a financial hub, I don’t think they would establish the Great Firewall [there],” says Chung Ching Kwong, a senior analyst at the Inter-Parliamentary Alliance on China, an advocacy organization that connects legislators from over 30 countries working on relations with China. 

It’s also the reason why the Hong Kong government has recently come out to say that it won’t outright ban platforms like Telegram and Signal, even though it said that it had received comments from the public asking it to do so.

But coming back to the court decision to restrict “Glory to Hong Kong,” even if the government doesn’t end up enforcing a full-blown ban of the song, as opposed to the more targeted injunction it’s imposed now, it may still result in significant harm to internet freedom.

We are still watching the responses roll in after the court decision last Wednesday. The Hong Kong government is anxiously waiting to hear how Google will react. Meanwhile, some videos have already been taken down, though it’s unclear whether they were pulled by the creators or by the platform. 

Michael Mo, a former district councilor in Hong Kong who’s now a postgraduate researcher at the University of Leeds in the UK, created a website right after the injunction was first initiated last June to embed all but one of the YouTube videos the government sought to ban. 

The domain name, “gloryto.hk,” was the first test of whether the Hong Kong domain registry would have trouble with it, but nothing has happened to it so far. The second test was seeing how soon the videos would be taken down on YouTube, which is now easy to tell by how many “video unavailable” gaps there are on the page. “Those videos were pretty much intact until the Court of Appeal overturned the rulings of the High Court. The first two have gone,” Mo says. 

The court case is having a chilling effect. Even entities that are not governed by the Hong Kong court are taking precautions. Some YouTube accounts owned by media based in Taiwan and the US proactively enabled geo-blocking to restrict people in Hong Kong from watching clips of the song they uploaded as soon as the injunction application was filed, Mo says. 

Are you optimistic or pessimistic about the future of internet freedom in Hong Kong? Let me know what you think at zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. The Biden administration plans to raise tariffs on Chinese-made EVs, from 25% to 100%. Since few Chinese cars are currently sold in the US, this is mostly a move to deter future imports of Chinese EVs. But it could slow down the decarbonization timeline in the US.  (ABC News)

2. Government officials from the US and China met in Geneva today to discuss how to mitigate the risks of AI. It’s a notable event, given how rare it is for the two sides to find common ground in the highly politicized field of technology. (Reuters $)

3. It will be more expensive soon to ride the bullet trains in China. A 20% to 39% fare increase is causing controversy among Chinese people. (New York Times $)

4. From executive leadership to workplace culture, TikTok has more in common with its Chinese sister app Douyin than the company wants to admit. (Rest of World)

5. China’s most indebted local governments have started claiming troves of data as “intangible assets” on their accounting books. Given the insatiable appetite for AI training data, they may have a point. (South China Morning Post $)

6. A crypto company with Chinese roots purchased a piece of land in Wyoming for crypto mining. Now the Biden administration is blocking the deal for national security reasons. (Associated Press)

Lost in translation

Recently, following an order made by the government, hotels in many major Chinese cities stopped asking guests to submit to facial recognition during check-in. 

According to the Chinese publication TechSina, this has had a devastating impact on the industry of facial recognition hardware. 

As hotels around the country retire their facial recognition kiosks en masse, equipment made by major tech companies has flooded online secondhand markets at steep discounts. What was sold for thousands of dollars is now resold for as little as 1% of the original price. Alipay, the Alibaba-affiliated payment app, once invested hundreds of millions of dollars to research and roll out these kiosks. Now it’s one of the companies being hit the hardest by the policy change.

One more thing

I had to double-check that this is not a joke. It turns out that for the past 10 years, the Louvre museum has been giving visitors a Nintendo 3DS—a popular handheld gaming console—as an audio and visual guide. 

It feels weird seeing people holding a 3DS up to the Mona Lisa as if they were in their own private Pokémon Go–style gaming world rather than just enjoying the museum. But apparently it doesn’t work very well anyway. Oops.

and it was THE WORST at navigating bc a 3ds can’t tell which direction you’re facing + the floorplan isn’t updated to match ongoing renovations. kept tryna send me into a wall i almost chucked the thing i stg

— taylor (@taylorhansss) May 12, 2024

It wasn’t exactly surprising when on Wednesday, May 8, a Hong Kong appeals court sided with the city government to take down “Glory to Hong Kong” from the internet. The trial, in which no one represented the defense, was the culmination of a years-long battle over a song that has become the unofficial anthem for protesters fighting China’s tightening control and police brutality in the city. But it remains an open question how exactly Big Tech will respond. Even as the injunction is narrowly designed to make it easier for them to comply, these Western companies may be seen as aiding authoritarian control and obstructing internet freedom if they do so.  

Google, Apple, Meta, Spotify, and others have spent the last several years largely refusing to cooperate with previous efforts by the Hong Kong government to prevent the spread of the song, which the government has claimed is a threat to national security. But the government has also hesitated to leverage criminal law to force them to comply with requests for removal of content, which could risk international uproar and hurt the city’s economy. 

Now, the new ruling seemingly finds a third option: imposing a civil injunction that doesn’t invoke criminal prosecution, which is similar to how copyright violations are enforced. Theoretically, the platforms may face less reputational blowback when they comply with this court order.

“If you look closely at the judgment, it’s basically tailor-made for the tech companies at stake,” says Chung Ching Kwong, a senior analyst at the Inter-Parliamentary Alliance on China, an advocacy organization that connects legislators from over 30 countries working on relations with China. She believes the language in the judgment suggests the tech companies will now be ready to comply with the government’s request.

A Google spokesperson said the company is reviewing the court’s judgment and didn’t respond to specific questions sent by MIT Technology Review. A Meta spokesperson pointed to a statement from Jeff Paine, the managing director of the Asia Internet Coalition, a trade group representing many tech companies in the Asia-Pacific region: “[The AIC] is assessing the implications of the decision made today, including how the injunction will be implemented, to determine its impact on businesses. We believe that a free and open internet is fundamental to the city’s ambitions to become an international technology and innovation hub.” The AIC did not immediately reply to questions sent via email. Apple and Spotify didn’t immediately respond to requests for comment.

But no matter what these companies do next, the ruling is already having an effect. Just over 24 hours after the court order, some of the 32 YouTube videos that are explicitly targeted in the injunction were inaccessible for users worldwide, not just in Hong Kong. 

While it’s unclear whether the videos were removed by the platform or by their creators, experts say the court decision will almost certainly set a precedent for more content to be censored from Hong Kong’s internet in the future.

“Censorship of the song would be a clear violation of internet freedom and freedom of expression,” says Yaqiu Wang, the research director for China, Hong Kong, and Taiwan at Freedom House, a human rights advocacy group. “Google and other internet companies should use all available channels to challenge the decision.” 

Erasing a song from the internet

Since “Glory to Hong Kong” was first uploaded to YouTube in August 2019 by an anonymous group called Dgx Music, it’s been adored by protesters and applauded as their anthem. Its popularity only grew after China passed the harsh Hong Kong national security law in 2020. 

With lyrics like “Liberate Hong Kong, revolution of our times,” it’s no surprise that it became a major flash point. The city and national Chinese governments were wary of its spread. 

Their fears escalated when the song was repeatedly mistaken for China’s national anthem at international events and was broadcast at sporting events after Hong Kong athletes won. By mid-2023 the mistake, intentional or not, had happened 887 times, according to the Hong Kong government’s request for the content’s removal, which cites YouTube videos and Google search results referring to the song as the “Hong Kong National Anthem” as the reason. 

The government has been arresting people for performing the song on the ground in Hong Kong, but it has been harder to prosecute the online activity since most of the videos and music were uploaded anonymously, and Hong Kong, unlike mainland China, has historically had a free internet. This meant officials needed to explore new approaches to content removal. 

To comply or not to comply

Using the controversial 2020 national security law as legal justification to make requests for removal of certain content that it deems threatening, the Hong Kong government has been able to exert pressure on local companies, like internet service providers. “In Hong Kong, all the major internet service providers are locally owned or Chinese-owned. For business reasons, probably within the last 20 years, most of the foreign investors like Verizon left on their own,” says Charles Mok, a researcher at Stanford University’s Cyber Policy Center and a former legislator in Hong Kong. “So right now, the government is focusing on telling the customer-facing internet service providers to do the blocking.” And it seems to have been somewhat effective, with a few websites for human rights organizations becoming inaccessible locally.

But the city government can’t get its way as easily when the content is on foreign-owned platforms like YouTube or Facebook. Back in 2020, most major Western companies declared they would pause processing data requests from the Hong Kong government while they assessed the law. Over time, some of them have started answering government requests again. But they’ve largely remained firm: over the first six months of 2023, for example, Meta received 41 requests from the Hong Kong government to obtain user data and answered none; during the same period, Google received requests to remove 164 items from Google services and ended up removing 82 of them, according to both companies’ transparency reports. Google specifically mentioned that it chose to not remove two YouTube videos and one Google Drive file related to “Glory to Hong Kong.”

Both sides are in tight spots. Tech companies don’t want to lose the Hong Kong market or endanger their local staff, but they are also worried about being seen as complying with authoritarian government actions. And the Hong Kong government doesn’t want to be seen as openly fighting Western platforms while trust in the region’s financial markets is already in decline. In particular, officials fear international headlines if the government invokes criminal law to force tech companies to remove certain content. 

“I think both sides are navigating this balancing act. So the government finally figured out a way that they thought might be able to solve the impasse: by going to the court and narrowly seeking an injunction,” Mok says.

That happened in June 2023, when Hong Kong’s government requested a court injunction to ban the distribution of the song online with the purpose of “inciting others to commit secession.” It named 32 YouTube videos explicitly, including the original version and live performances, translations into other languages, instrumental and opera versions, and an interview with the original creators. But the order would also cover “any adaptation of the song, the melody and/or lyrics of which are substantially the same as the song,” according to court documents. 

The injunction went through a year of back-and-forth hearings, including a lower court ruling that briefly swatted down the ban. But now, the Court of Appeal has granted the government approval. The case can theoretically be appealed one last time, but with no defendants present, that’s unlikely to happen.

The key difference between this action and previous attempts to remove content is that this is a civil injunction, not a criminal prosecution—meaning it is, at least legally speaking, closer to a copyright takedown request. A platform could arguably be less likely to take a reputational hit if it removes the content upon request. 

Kwong believes this will indeed make platforms more likely to cooperate, and there have already been pretty clear signs to that effect. In one hearing in December, the government was asked by the court to consult online platforms as to the feasibility of the injunction. The final judgment this week says that while the platforms “have not taken part in these proceedings, they have indicated that they are ready to accede to the Government’s request if there is a court order.”

“The actual targets in this case, mainly the tech giants, may have less hesitation to comply with a civil court order than a national security order because if it’s the latter, they may also face backfire from the US,” says Eric Yan-Ho Lai, a research fellow at Georgetown Center for Asian Law. 

Lai also says now that the injunction is granted, it will be easier to prosecute an individual based on violation of a civil injunction rather than prosecuting someone for criminal offenses, since the government won’t need to prove criminal intent.

The chilling effect

Immediately after the injunction, human rights advocates called on tech companies to remain committed to their values. “Companies like Google and Apple have repeatedly claimed that they stand by the universal right to freedom of expression. They should put their ideals into practice,” says Freedom House’s Wang. “Google and other tech companies should thoroughly document government demands, and publish detailed transparency reports on content takedowns, both for those initiated by the authorities and those done by the companies themselves.”

Without making their plans clear, it’s too early to know just how tech companies will react. But right after the injunction was granted, the song largely remained available for Hong Kong users on most platforms, including YouTube, iTunes, and Spotify, according to the South China Morning Post. On iTunes, the song even returned to the top of the download rankings a few hours after the injunction.

One key factor that may still determine corporate cooperation is how far the content removal requests go. There will surely be more videos of the song that are uploaded to YouTube, not to mention independent websites hosting the videos and music for more people to access. Will the government go after each of them too?

The Hong Kong government has previously said in court hearings that it seeks only local restriction of the online content, meaning content will be inaccessible only to users physically in the city. Large platforms like YouTube can do that without difficulty. 

Theoretically, this allows local residents to circumvent the ban by using VPN software, but not everyone is technologically savvy enough to do so. And that wouldn’t do much to minimize the larger chilling effect on free speech, says Kwong from the Inter-Parliamentary Alliance on China. 

“As a Hong Konger living abroad, I do rely on Hong Kong services or international services based in Hong Kong to get ahold of what’s happening in the city. I do use YouTube Hong Kong to see certain things, and I do use Spotify Hong Kong or Apple Music because I want access to Cantopop,” she says. “At the same time, you worry about what you can share with friends in Hong Kong and whatnot. We don’t want to put them into trouble by sharing things that they are not supposed to see, which they should be able to see.”

The court made at least two explicit exemptions to the song’s ban, for “lawful activities conducted in connection with the song, such as those for the purpose of academic activity and news activity.” But even the implementation of these could be incredibly complex and confusing in practice. “In the current political context in Hong Kong, I don’t see anyone willing to take the risk,” Kwong says. 

The government has already arrested prominent journalists on accusations of endangering national security, and a new law passed in 2024 has expanded the crimes that can be prosecuted on national security grounds. As with all efforts to suppress free speech, the impact of vague boundaries that encourage self-censorship on potentially sensitive topics is often sprawling and hard to measure. 

“Nobody knows where the actual red line is,” Kwong says.

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

If you could talk again to someone you love who has passed away, would you? For a long time, this has been a hypothetical question. No longer. 

Deepfake technologies have evolved to the point where it’s now easy and affordable to clone people’s looks and voices with AI. Meanwhile, large language models mean it’s more feasible than ever before to conduct full conversations with AI chatbots. 

I just published a story today about the burgeoning market in China for applying these advances to re-create deceased family members. Thousands of grieving individuals have started turning to dead relatives’ digital avatars for conversations and comfort. 

It’s a modern twist on a cultural tradition of talking to the dead, whether at their tombs, during funeral rituals, or in front of their memorial portraits. Chinese people have always liked to tell lost loved ones what has happened since they passed away. But what if the dead could talk back? This is the proposition of at least half a dozen Chinese companies offering “AI resurrection” services. The products, costing a few hundred to a few thousand dollars, are lifelike avatars, accessed in an app or on a tablet, that let people interact with the dead as if they were still alive.

I talked to two Chinese companies that, combined, have provided this service for over 2,000 clients. They describe a growing market of people accepting the technology. Their customers usually look to the products to help them process their grief.

To read more about how these products work and the potential implications of the technology, go here.

However, what I didn’t get into in the story is that the same technology used to clone the dead has also been used in other interesting ways.

For one, this process is being applied not just to private individuals, but also to public figures. Sima Huapeng, CEO and cofounder of the Chinese company Silicon Intelligence, tells me that about one-third of the “AI resurrection” cases he has worked on involve making avatars of dead Chinese writers, thinkers, celebrities, and religious leaders. The generated product is not intended for personal mourning but more for public education or memorial purposes.

Last year, Silicon Intelligence replicated Mei Lanfang, a renowned Peking opera singer born in 1894. The avatar of Mei was commissioned to address a 2023 Peking opera festival held in his hometown, Taizhou. Mei talked about seeing how drastically Taizhou had changed through modern urban development, even though the real artist died in 1961.

But an even more interesting use of this technology is that people are using it to clone themselves while they are still alive, to preserve their memories and leave a legacy. 

Sima said this is becoming more popular among successful families that feel the need to pass on their stories. He showed me a video of an avatar the company created for a 92-year-old Chinese entrepreneur, which was displayed on a big vertical monitor screen. The entrepreneur wrote a book documenting his life, and the company only had to feed the whole book to a large language model for it to start role-playing him. “This grandpa cloned himself so he could pass on the stories of his life to the whole family. Even when he dies, he can still talk to his descendants like this,” says Sima.

Sun Kai, another cofounder of Silicon Intelligence, is also featured in my story because he made a replica of his mom, who passed away in 2019. One of his regrets is that he didn’t have enough video recordings of his mom that he could use to train her avatar to be more like her. That inspired him to start recording voice memos of his life and working on his own digital “twin,” even though, in his 40s, death still seems far away.

He compares the process to a complicated version of a photo shoot, but a digital avatar that has his looks, voice, and knowledge can preserve much more information than photographs do. 

And there’s still another use: Just as parents can spend money on an expensive photo shoot to capture their children at a specific age, they can also choose to create an AI avatar for the same purpose. “The parents tell us no matter how many photos or videos they took of their 12-year-old kid, it always felt like something was lacking. But once we digitized this kid, they could talk to the 12-year-old version of them anytime, anywhere,” Sun says.

At the end of the day, the deepfake technologies used to clone both the living and the deceased are the same. And seeing that there’s already a market in China for such services, I’m sure these companies will keep on developing more use cases for it. 

But what’s also certain is that we’d have to answer a lot more questions about the ethical challenges of these applications, from the issue of consent to violations of copyright. 

Would you make a replica of yourself if given the chance? Tell me your thoughts at zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. Zhang Yongzhen, the first Chinese scientist to publish a sequence of the covid-19 virus, staged a protest last week over being locked out of his lab—likely a result of the Chinese government’s efforts to discourage research on covid origins. (Associated Press $)

2. Chinese president Xi Jinping is visiting Europe for five days. Half of the trip will be spent in Hungary and Serbia, the only two European countries that are welcoming Chinese investment and manufacturing. Xi is expected to announce an electric-vehicle manufacturing deal in Hungary while he’s there. (Associated Press)

3. China launched a new moon-exploring rover on Friday. It will collect samples near the moon’s south pole, an area where the US and China are competing to build permanent bases. Maybe the Netflix comedy series Space Force will look like a documentary soon. (Wall Street Journal $)

4. Huawei is secretly funding an optics research competition in the US. The act likely isn’t illegal, but it’s deceptive, since university participants, some of whom had vowed to not work with the company, didn’t know the source of the funding. (Bloomberg $)

5. China is quickly catching up on brain-computer interfaces, and there’s strong interest in using the technology for non-medical cognitive improvement. (Wired $)

6. Taiwan has been rocked by frequent earthquakes this year, and developers are racing to make earthquake warning apps that might save lives. One such app has seen user numbers increase from 3,000 to 370,000. (Reuters $)

7. Prestigious Chinese media publications, which still publish hard-hitting stories at times, are being forced to distance themselves from the highest-profile journalism award in Asia to avoid being accused by the government of “colluding with foreign forces.”  (Nikkei Asia $)

Lost in translation

While generative AI companies have taken the spotlight during the current AI frenzy, China’s older “AI Four Dragons”—four companies that rose to market prominence because of their technological lead in computer vision and facial recognition—are grappling with profit setbacks and commercialization hurdles, reports the Chinese publication Guiji Yanjiushi.

In response to these challenges, the “Dragons” have chosen different strategies. Yitu leaned further into security cameras; Megvii focused on applying computer vision in logistics and the Internet of Things; CloudWalk prioritized AI assistants; and SenseTime, the largest of them all, ventured into generative AI with its self-developed LLMs. Even though they are not as trendy as the startups, some experts believe these established players, having accumulated more computing power and AI talent over the years, may prove to be more resilient in the end.

One more thing

During this year’s Met Gala, fans were struggling to discern real photos of celebrities from AI-generated ones. To add to the confusion, some social media accounts were running real photos in AI-powered enhancement apps, which slightly distorted the images and made it even harder to tell the difference. 

One of the most widely used such apps is called Remini, but few people know that it was actually developed by a Chinese company called Caldron and later acquired by an Italian software company. Remini now has over 20 million users and is extremely profitable. Still, it seems its AI enhancement tools have a long way to go.

bestie… @2015smetgala it’s time to delete the remini app… you’ve gone too far https://t.co/Q4Aj2454U8 pic.twitter.com/yqH46EJlJd

— swiftie wins (@swifferwins) May 7, 2024

Once a week, Sun Kai has a video call with his mother. He opens up about work, the pressures he faces as a middle-aged man, and thoughts that he doesn’t even discuss with his wife. His mother will occasionally make a comment, like telling him to take care of himself—he’s her only child. But mostly, she just listens.

That’s because Sun’s mother died five years ago. And the person he’s talking to isn’t actually a person, but a digital replica he made of her—a moving image that can conduct basic conversations. They’ve been talking for a few years now. 

After she died of a sudden illness in 2019, Sun wanted to find a way to keep their connection alive. So he turned to a team at Silicon Intelligence, an AI company based in Nanjing, China, that he cofounded in 2017. He provided them with a photo of her and some audio clips from their WeChat conversations. While the company was mostly focused on audio generation, the staff spent four months researching synthetic tools and generated an avatar with the data Sun provided. Then he was able to see and talk to a digital version of his mom via an app on his phone. 

“My mom didn’t seem very natural, but I still heard the words that she often said: ‘Have you eaten yet?’” Sun recalls of the first interaction. Because generative AI was a nascent technology at the time, the replica of his mom can say only a few pre-written lines. But Sun says that’s what she was like anyway. “She would always repeat those questions over and over again, and it made me very emotional when I heard it,” he says.

There are plenty of people like Sun who want to use AI to preserve, animate, and interact with lost loved ones as they mourn and try to heal. The market is particularly strong in China, where at least half a dozen companies are now offering such technologies and thousands of people have already paid for them. In fact, the avatars are the newest manifestation of a cultural tradition: Chinese people have always taken solace from confiding in the dead. 

The technology isn’t perfect—avatars can still be stiff and robotic—but it’s maturing, and more tools are becoming available through more companies. In turn, the price of “resurrecting” someone—also called creating “digital immortality” in the Chinese industry—has dropped significantly. Now this technology is becoming accessible to the general public. 

Some people question whether interacting with AI replicas of the dead is actually a healthy way to process grief, and it’s not entirely clear what the legal and ethical implications of this technology may be. For now, the idea still makes a lot of people uncomfortable. But as Silicon Intelligence’s other cofounder, CEO Sima Huapeng, says, “Even if only 1% of Chinese people can accept [AI cloning of the dead], that’s still a huge market.” 

AI resurrection

Avatars of the dead are essentially deepfakes: the technologies used to replicate a living person and a dead person aren’t inherently different. Diffusion models generate a realistic avatar that can move and speak. Large language models can be attached to generate conversations. The more data these models ingest about someone’s life—including photos, videos, audio recordings, and texts—the more closely the result will mimic that person, whether dead or alive.

China has proved to be a ripe market for all kinds of digital doubles. For example, the country has a robust e-commerce sector, and consumer brands hire many livestreamers to sell products. Initially, these were real people—but as MIT Technology Review reported last fall—many brands are switching to AI-cloned influencers that can stream 24/7. 

In just the past three years, the Chinese sector developing AI avatars has matured rapidly, says Shen Yang, a professor studying AI and media at Tsinghua University in Beijing, and replicas have improved from minutes-long rendered videos to 3D “live” avatars that can interact with people.  

This year, Sima says, has seen a tipping point, with AI cloning becoming affordable for most individuals. “Last year, it cost about $2,000 to $3,000, but it now only costs a few hundred dollars,” he says. That’s thanks to a price war between Chinese AI companies, which are fighting to meet the thriving demand for digital avatars in other sectors like streaming.

In fact, demand for applications that re-create the dead has also boosted the capabilities of tools that digitally replicate the living. 

Silicon Intelligence offers both services. When Sun and Sima launched the company, they were focused on using text-to-speech technologies to create audio and then using those AI-generated voices in applications such as robocalls.

But after the company replicated Sun’s mother, it pivoted to generating realistic avatars. That decision turned the company into one of the leading Chinese players creating AI-powered influencers. 

SILICON INTELLIGENCE

Its technology has generated avatars for hundreds of thousands of TikTok-like videos and streaming channels, but Sima says more recently it’s seen around 1,000 clients use it to replicate someone who’s passed away. “We started our work on ‘resurrection’ in 2019 and 2020,” he says, but at first people were slow to accept it: “No one wanted to be the first adopters.” 

The quality of the avatars has improved, he says, which has boosted adoption. When the avatar looks increasingly lifelike and gives fewer out-of-character answers, it’s easier for users to treat it as their deceased family member. Plus, the idea is getting popularized through more depictions on Chinese TV. 

Now Silicon Intelligence offers the replication service for a price between several hundred and several thousand dollars. The most basic product comes as an interactive avatar in an app, and the options at the upper end of the range often involve more customization and better hardware components, such as a tablet or a display screen. There are at least a handful more Chinese companies working on the same technology.

A modern twist on tradition

The business in these deepfakes builds on China’s long cultural history of communicating with the dead. 

In Chinese homes, it’s common to put up a portrait of a deceased relative for a few years after the death. Zhang Zewei, founder of a Shanghai-based company called Super Brain, says he and his team wanted to revamp that tradition with an “AI photo frame.” They create avatars of deceased loved ones that are pre-loaded onto an Android tablet, which looks like a photo frame when standing up. Clients can choose a moving image that speaks words drawn from an offline database or from an LLM. 

“In its essence, it’s not much different from a traditional portrait, except that it’s interactive,” Zhang says.

Zhang says the company has made digital replicas for over 1,000 clients since March 2023 and charges $700 to $1,400, depending on the service purchased. The company plans to release an app-only product soon, so that users can access the avatars on their phones, and could further reduce the cost to around $140.

SUPER BRAIN

The purpose of his products, Zhang says, is therapeutic. “When you really miss someone or need consolation during certain holidays, you can talk to the artificial living and heal your inner wounds,” he says.

And even if that conversation is largely one-sided, that’s in keeping with a strong cultural tradition. Every April during the Qingming festival, Chinese people sweep the tombs of their ancestors, burn joss sticks and fake paper money, and tell them what has happened in the past year. Of course, those conversations have always been one-way. 

But that’s not the case for all Super Brain services. The company also offers deepfaked video calls in which a company employee or a contract therapist pretends to be the relative who passed away. Using DeepFace, an open-source tool that analyzes facial features, the deceased person’s face is reconstructed in 3D and swapped in for the live person’s face with a real-time filter. 

At the other end of the call is usually an elderly family member who may not know that the relative has died—and whose family has arranged the conversation as a ruse. 

Jonathan Yang, a Nanjing resident who works in the tech industry, paid for this service in September 2023. His uncle died in a construction accident, but the family hesitated to tell Yang’s grandmother, who is 93 and in poor health. They worried that she wouldn’t survive the devastating news.

So Yang paid $1,350 to commission three deepfaked calls of his dead uncle. He gave Super Brain a handful of photos and videos of his uncle to train the model. Then, on three Chinese holidays, a Super Brain employee video-called Yang’s grandmother and told her, as his uncle, that he was busy working in a faraway city and wouldn’t be able to come back home, even during the Chinese New Year. 

“The effect has met my expectations. My grandma didn’t suspect anything,” Yang says. His family did have mixed opinions about the idea, because some relatives thought maybe she would have wanted to see her son’s body before it was cremated. Still, the whole family got on board in the end, believing the ruse would be best for her health. After all, it’s pretty common for Chinese families to tell “necessary” lies to avoid overwhelming seniors, as depicted in the movie The Farewell. 

To Yang, a close follower of the AI industry trends, creating replicas of the dead is one of the best applications of the technology. “It best represents the warmth [of AI],” he says. His grandmother’s health has improved, and there may come a day when they finally tell her the truth. By that time, Yang says, he may purchase a digital avatar of his uncle for his grandma to talk to whenever she misses him.

Is AI really good for grief? 

Even as AI cloning technology improves, there are some significant barriers preventing more people from using it to speak with their dead relatives in China. 

On the tech side, there are limitations to what AI models can generate. Most LLMs can handle dominant languages like Mandarin and Cantonese, but they aren’t able to replicate the many niche dialects in China. It’s also challenging—and therefore costly—to replicate body movements and complex facial expressions in 3D models. 

Then there’s the issue of training data. Unlike cloning someone who’s still alive, which often involves asking the person to record body movements or say certain things, posthumous AI replications must rely on whatever videos or photos are already available. And many clients don’t have high-quality data, or enough of it, for the end result to be satisfactory. 

Complicating these technical challenges are myriad ethical questions. Notably, how can someone who is already dead consent to being digitally replicated? For now, companies like Super Brain and Silicon Intelligence rely on the permission of direct family members. But what if family members disagree? And if a digital avatar generates inappropriate answers, who is responsible?

Similar technology caused controversy earlier this year. A company in Ningbo reportedly used AI tools to create videos of deceased celebrities and posted them on social media to speak to their fans. The videos were generated using public data, but without seeking any approval or permission. The result was intense criticism from the celebrities’ families and fans, and the videos were eventually taken down. 

“It’s a new domain that only came about after the popularization of AI: the rights to digital eternity,” says Shen, the Tsinghua professor, who also runs a lab that creates digital replicas of people who have passed away. He believes it should be prohibited to use deepfake technology to replicate living people without their permission. For people who have passed away, all of their immediate living family members must agree beforehand, he says. 

There could be negative effects on clients’ mental health, too. While some people, like Sun, find their conversations with avatars to be therapeutic, not everyone thinks it’s a healthy way to grieve. “The controversy lies in the fact that if we replicate our family members because we miss them, we may constantly stay in the state of mourning and can’t withdraw from it to accept that they have truly passed away,” says Shen. A widowed person who’s in constant conversation with the digital version of their partner might be held back from seeking a new relationship, for instance. 

“When someone passes away, should we replace our real emotions with fictional ones and linger in that emotional state?” Shen asks. Psychologists and philosophers who talked to MIT Technology Review about the impact of grief tech have warned about the danger of doing so. 

Sun Kai, at least, has found the digital avatar of his mom to be a comfort. She’s like a 24/7 confidante on his phone. Even though it’s possible to remake his mother’s avatar with the latest technology, he hasn’t yet done that. “I’m so used to what she looks like and sounds like now,” he says. As years have gone by, the boundary between her avatar and his memory of her has begun to blur. “Sometimes I couldn’t even tell which one is the real her,” he says.

And Sun is still okay with doing most of the talking. “When I’m confiding in her, I’m merely letting off steam. Sometimes you already know the answer to your question, but you still need to say it out loud,” he says. “My conversations with my mom have always been like this throughout the years.” 

But now, unlike before, he gets to talk to her whenever he wants to.

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

Allow me to indulge in a little reflection this week. Last week, the divest-or-ban TikTok bill was passed in Congress and signed into law. Four years ago, when I was just starting to report on the world of Chinese technologies, one of my first stories was about very similar news: President Donald Trump announcing he’d ban TikTok. 

That 2020 executive order came to nothing in the end—it was blocked in the courts, put aside after the presidency changed hands, and eventually withdrawn by the Biden administration. Yet the idea—that the US government should ban TikTok in some way—never went away. It would repeatedly be suggested in different forms and shapes. And eventually, on April 24, 2024, things came full circle.

A lot has changed in the four years between these two news cycles. Back then, TikTok was a rising sensation that many people didn’t understand; now, it’s one of the biggest social media platforms, the originator of a generation-defining content medium, and a music-industry juggernaut. 

What has also changed is my outlook on the issue. For a long time, I thought TikTok would find a way out of the political tensions, but I’m increasingly pessimistic about its future. And I have even less hope for other Chinese tech companies trying to go global. If the TikTok saga tells us anything, it’s that their Chinese roots will be scrutinized forever, no matter what they do.

I don’t believe TikTok has become a larger security threat now than it was in 2020. There have always been issues with the app, like potential operational influence by the Chinese government, the black-box algorithms that produce unpredictable results, and the fact that parent company ByteDance never managed to separate the US side and the China side cleanly, despite efforts (one called Project Texas) to store and process American data locally. 

But none of those problems got worse over the last four years. And interestingly, while discussions in 2020 still revolved around potential remedies like setting up data centers in the US to store American data or having an organization like Oracle audit operations, those kinds of fixes are not in the law passed this year. As long as it still has Chinese owners, the app is not permissible in the US. The only thing it can do to survive here is transfer ownership to a US entity. 

That’s the cold, hard truth not only for TikTok but for other Chinese companies too. In today’s political climate, any association with China and the Chinese government is seen as unacceptable. It’s a far cry from the 2010s, when Chinese companies could dream about developing a killer app and finding audiences and investors around the globe—something many did pull off. 

There’s something I wrote four years ago that still rings true today: TikTok is the bellwether for Chinese companies trying to go global. 

The majority of Chinese tech giants, like Alibaba, Tencent, and Baidu, operate primarily within China’s borders. TikTok was the first to gain mass popularity in lots of other countries across the world and become part of daily life for people outside China. To many Chinese startups, it showed that the hard work of trying to learn about foreign countries and users can eventually pay off, and it’s worth the time and investment to try.

On the other hand, if even TikTok can’t get itself out of trouble, with all the resources that ByteDance has, is there any hope for the smaller players?

When TikTok found itself in trouble, the initial reaction of these other Chinese companies was to conceal their roots, hoping they could avoid attention. During my reporting, I’ve encountered multiple companies that fret about being described as Chinese. “We are headquartered in Boston,” one would say, while everyone in China openly talked about its product as the overseas version of a Chinese app.

But with all the political back-and-forth about TikTok, I think these companies are also realizing that concealing their Chinese associations doesn’t work—and it may make them look even worse if it leaves users and regulators feeling deceived.

With the new divest-or-ban bill, I think these companies are getting a clear signal that it’s not the technical details that matter—only their national origin. The same worry is spreading to many other industries, as I wrote in this newsletter last week. Even in the climate and renewable power industries, the presence of Chinese companies is becoming increasingly politicized. They, too, are finding themselves scrutinized more for their Chinese roots than for the actual products they offer.

Obviously, none of this is good news to me. When they feel unwelcome in the US market, Chinese companies don’t feel the need to talk to international media anymore. Without these vital conversations, it’s even harder for people in other countries to figure out what’s going on with tech in China.

Instead of banning TikTok because it’s Chinese, maybe we should go back to focus on what TikTok did wrong: why certain sensitive political topics seem deprioritized on the platform; why Project Texas has stalled; how to make the algorithmic workings of the platform more transparent. These issues, instead of whether TikTok is still controlled by China, are the things that actually matter. It’s a harder path to take than just banning the app entirely, but I think it’s the right one.

Do you believe the TikTok ban will go through? Let me know your thoughts at zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. Facing the possibility of a total ban on TikTok, influencers and creators are making contingency plans. (Wired $)

2. TSMC has brought hundreds of Taiwanese employees to Arizona to build its new chip factory. But the company is struggling to bridge cultural and professional differences between American and Taiwanese workers. (Rest of World)

3. The US secretary of state, Antony Blinken, met with Chinese president Xi Jinping during a visit to China this week. (New York Times $)

• Here’s the best way to describe these recent US-China diplomatic meetings: “The US and China talk past each other on most issues, but at least they’re still talking.” (Associated Press)

4. Half of Russian companies’ payments to China are made through middlemen in Hong Kong, Central Asia, or the Middle East to evade sanctions. (Reuters $)

5. A massive auto show is taking place in Beijing this week, with domestic electric vehicles unsurprisingly taking center stage. (Associated Press)

• Meanwhile, Elon Musk squeezed in a quick trip to China and met with his “old friend” the Chinese premier Li Qiang, who was believed to have facilitated establishing the Gigafactory in Shanghai. (BBC)
• Tesla may finally get a license to deploy its autopilot system, which it calls Full Self Driving, in China after agreeing to collaborate with Baidu. (Reuters $)

6. Beijing has hosted two rival Palestinian political groups, Hamas and Fatah, to talk about potential reconciliation. (Al Jazeera)

Lost in translation

The Chinese dubbing community is grappling with the impacts of new audio-generating AI tools. According to the Chinese publication ACGx, for a new audio drama, a music company licensed the voice of the famous dubbing actor Zhao Qianjing and used AI to transform it into multiple characters and voice the entire script. 

But online, this wasn’t really celebrated as an advancement for the industry. Beyond criticizing the quality of the audio drama (saying it still doesn’t sound like real humans), dubbers are worried about the replacement of human actors and increasingly limited opportunities for newcomers. Other than this new audio drama, there have been several examples in China where AI audio generation has been used to replace human dubbers in documentaries and games. E-book platforms have also allowed users to choose different audio-generated voices to read out the text. 

One more thing

While in Beijing, Antony Blinken visited a record store and bought two vinyl records—one by Taylor Swift and another by the Chinese rock star Dou Wei. Many Chinese (and American!) people learned for the first time that Blinken had previously been in a rock band.

Almost all keyboard apps used by Chinese people around the world share a security loophole that makes it possible to spy on what users are typing. 

The vulnerability, which allows the keystroke data that these apps send to the cloud to be intercepted, has existed for years and could have been exploited by cybercriminals and state surveillance groups, according to researchers at the Citizen Lab, a technology and security research lab affiliated with the University of Toronto.

These apps help users type Chinese characters more efficiently and are ubiquitous on devices used by Chinese people. The four most popular apps—built by major internet companies like Baidu, Tencent, and iFlytek—basically account for all the typing methods that Chinese people use. Researchers also looked into the keyboard apps that come preinstalled on Android phones sold in China. 

What they discovered was shocking. Almost every third-party app and every Android phone with preinstalled keyboards failed to protect users by properly encrypting the content they typed. A smartphone made by Huawei was the only device where no such security vulnerability was found.

In August 2023, the same researchers found that Sogou, one of the most popular keyboard apps, did not use Transport Layer Security (TLS) when transmitting keystroke data to its cloud server for better typing predictions. Without TLS, a widely adopted international cryptographic protocol that protects users from a known encryption loophole, keystrokes can be collected and then decrypted by third parties.

“Because we had so much luck looking at this one, we figured maybe this generalizes to the others, and they suffer from the same kinds of problems for the same reason that the one did,” says Jeffrey Knockel, a senior research associate at the Citizen Lab, “and as it turns out, we were unfortunately right.”

Even though Sogou fixed the issue after it was made public last year, some Sogou keyboards preinstalled on phones are not updated to the latest version, so they are still subject to eavesdropping. 

This new finding shows that the vulnerability is far more widespread than previously believed. 

“As someone who also has used these keyboards, this was absolutely horrifying,” says Mona Wang, a PhD student in computer science at Princeton University and a coauthor of the report. 

“The scale of this was really shocking to us,” says Wang. “And also, these are completely different manufacturers making very similar mistakes independently of one another, which is just absolutely shocking as well.”

The massive scale of the problem is compounded by the fact that these vulnerabilities aren’t hard to exploit. “You don’t need huge supercomputers crunching numbers to crack this. You don’t need to collect terabytes of data to crack it,” says Knockel. “If you’re just a person who wants to target another person on your Wi-Fi, you could do that once you understand the vulnerability.” 

The ease of exploiting the vulnerabilities and the huge payoff—knowing everything a person types, potentially including bank account passwords or confidential materials—suggest that it’s likely they have already been taken advantage of by hackers, the researchers say. But there’s no evidence of this, though state hackers working for Western governments targeted a similar loophole in a Chinese browser app in 2011.

Most of the loopholes found in this report are “so far behind modern best practices” that it’s very easy to decrypt what people are typing, says Jedidiah Crandall, an associate professor of security and cryptography at Arizona State University, who was consulted in the writing of this report. Because it doesn’t take much effort to decrypt the messages, this type of loophole can be a great target for large-scale surveillance of massive groups, he says.

After the researchers got in contact with companies that developed these keyboard apps, the majority of the loopholes were fixed. Samsung, whose self-developed app was also found to lack sufficient encryption, sent MIT Technology Review an emailed statement: “We were made aware of potential vulnerabilities and have issued patches to address these issues. As always, we recommend that all users keep their devices updated with the latest software to ensure the highest level of protection possible.”

But a few companies have been unresponsive, and the vulnerability still exists in some apps and phones, including QQ Pinyin and Baidu, as well as in any keyboard app that hasn’t been updated to the latest version. Baidu, Tencent, and iFlytek did not reply to press inquiries sent by MIT Technology Review.

One potential cause of the loopholes’ ubiquity is that most of these keyboard apps were developed in the 2000s, before the TLS protocol was commonly adopted in software development. Even though the apps have been through numerous rounds of updates since then, inertia could have prevented developers from adopting a safer alternative.

The report points out that language barriers and different tech ecosystems prevent English- and Chinese-speaking security researchers from sharing information that could fix issues like this more quickly. For example, because Google’s Play store is blocked in China, most Chinese apps are not available in Google Play, where Western researchers often go for apps to analyze. 

Sometimes all it takes is a little additional effort. After two emails about the issue to iFlytek were met with silence, the Citizen Lab researchers changed the email title to Chinese and added a one-line summary in Chinese to the English text. Just three days later, they received an email from iFlytek, saying that the problem had been resolved.

Update: The story has been updated to include Samsung’s statement.

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

I’ve wanted to learn more about the world of solar panels ever since I realized just how dominant Chinese companies have become in this field. Although much of the technology involved was invented in the US, today about 80% of the world’s solar manufacturing takes place in China. For some parts of the process, it’s responsible for even more: 97% of wafer manufacturing, for example. 

So I jumped at the opportunity to interview Shawn Qu, the founder and chairman of Canadian Solar, one of the largest and longest-standing solar manufacturing companies in the world, last week.

Qu’s company provides a useful lens on wider efforts by the US to reshape the global solar supply chain and bring more of it back to American shores. Although most of its production is still in China and Southeast Asia, it’s now building two factories in the US, spurred on by incentives in the Inflation Reduction Act. You can read my story here.

I met Qu in Cambridge, Massachusetts, where he was attending the Harvard College China Forum, a two-day annual conference that often draws a fair number of Chinese entrepreneurs. I also attended, hoping to meet representatives of Chinese tech companies there.

At the conference, I noticed three interesting things.

One, there was a glaring absence of Chinese consumer tech companies. With the exception of one US-based manager from TikTok, I didn’t see anyone from Alibaba, Baidu, Tencent, or ByteDance. 

These companies, with their large influence on Chinese people’s everyday lives, used to be the stars of discussions around China’s tech sector. If you had come to the Harvard conference before covid-19, you would have met plenty of people representing them, as well as the venture capitalists that funded their successes. You can get a sense just by reading past speaker lists: executives from Xiaomi, Ant Financial, Sogou, Sequoia China, and Hillhouse Capital. These are the equivalents of Mark Zuckerberg and Peter Thiel in China’s tech world.

But these companies have become much more low profile since then, for a couple of main reasons. First, they underwent a harsh domestic crackdown after the government decided to tame them. (I recently talked to Angela Zhang, a law professor studying Chinese tech regulations, to understand these crackdowns.) And second, they have become the subject of national security scrutiny in the US, making it politically unwise for them to engage too much on the public stage here.

The second thing I noticed at the conference is what stood in their place: a batch of new Chinese companies, mostly in climate tech. William Li, the CEO of China’s EV startup NIO, was one of the most popular guest speakers during the conference’s opening ceremony this year. There were at least three solar panel companies present—two (JA Solar and Canadian Solar) among the top-tier manufacturers in the world, and a third that sells solar panels to Latin America. There were also many academics, investors, and even influencers working in the field of electric vehicles and other electrified transportation methods.

It’s clear that amid the increasingly urgent task of addressing climate change, China’s climate technology companies have become the new stars of the show. And they are very much willing to appear on the global stage, both bragging about their technological lead and seeking new markets. 

“The Chinese entrepreneurs are very eager,” says Jinhua Zhao, a professor studying urban transportation at MIT, who also spoke on one of the panels at the conference. “They want to come out. I think the Chinese government side also started to send signals, inviting foreign leadership and financial industries to visit China. I see a lot of gestures.” 

The problem, however, is they are also becoming subject to a lot of political animosity in the US. The Biden administration has started an investigation into Chinese-made cars, mostly electric vehicles; Chinese battery companies have been navigating a minefield of politicians’ resistance to their setting up plants in North America; and Chinese solar panel companies have been subject to sky-high tariffs. 

Back in the mid-2010s, when Chinese consumer tech companies emerged onto the global stage, the US and China had a warm relationship, creating a welcoming environment. Unfortunately, that’s not something climate tech companies can enjoy today. Even though climate change is a global issue that requires countries to collaborate, political tensions stand in the way when companies and investors on opposite sides try to work together.

On that note, the last thing I noticed at the conference is a rising geopolitical force in tech: the Middle East. A few speakers at the conference are working in Saudi Arabia and the United Arab Emirates, and they represent other deep-pocketed players who are betting on technologies like EVs and AI in both the United States and China.

But can they navigate the tensions and benefit from the technological advantages on both sides? It’ll be interesting to watch how that unfolds. 

What do you think of the role of the Middle East in the future of climate technologies? Let me know your thoughts at zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. A batch of documents mistakenly unsealed by a Pennsylvania court reveals the origin story of TikTok’s parent company, ByteDance. Who knew it started out as a real estate venture? (New York Times $)

2. Vladimir Potanin, Russia’s richest man, said he would move some of his copper smelting factories to China to reduce the impact of Western sanctions, which block Russian companies from using international payment systems. (Financial Times $)

3. Chinese universities have found a way to circumvent the US export ban on high-end Nvidia chips: by buying resold server products made by Dell, Super Micro Computer, and Taiwan’s Gigabyte Technology. (Reuters $)

4. TikTok is testing “TikTok Notes,” a rival product to Instagram, in Australia and Canada. (The Verge)

5. Since there’s no route for personal bankruptcy in China, those who are unable to pay their debts are being penalized in novel ways: they can’t take high-speed trains, fly on planes, stay in nice hotels, or buy expensive insurance policies. (Wall Street Journal $)

6. The hunt for the origins of covid-19 has stalled in China, as Chinese politicians worry about being blamed for the findings. (Associated Press)

7. Because of pressure from the US government, Mexico will not hand out tax cuts and other incentives to Chinese EV companies. (Reuters $)

Lost in translation

Until last year, it was normal for Chinese hotels to require facial recognition to check guests in, but the city of Shanghai is now turning against the practice, according to the Chinese publication 21st Century Business Herald. The police bureau of Shanghai recently published a notice that says “scanning faces” is required only if guests don’t have any identity documents. Otherwise, they have the right to refuse it. Most hotel chains in Shanghai, and some in other cities, have updated their policies in response. 

China has a national facial recognition database tied to the government ID system, and businesses such as hotels can access it to verify customers’ identities. However, Chinese people are increasingly pushing back on the necessity of facial recognition in scenarios like this, and questioning whether hotels are handling such sensitive biometric data properly. 

One more thing

The latest queer icon in Asia is Nymphia Wind, the drag persona of a 28-year-old Taiwanese-American named Leo Tsao, who just won the latest season of RuPaul’s Drag Race. Fully embracing the color yellow as part of her identity, Nymphia Wind is also called the “Banana Buddha” by her fans. She’s hosting shows in Taoist temples in Taiwan, attracting audiences old and young.

Whenever you see a solar panel, most parts of it probably come from China. The US invented the technology and once dominated its production, but over the past two decades, government subsidies and low costs in China have led most of the solar manufacturing supply chain to be concentrated there. The country will soon be responsible for over 80% of solar manufacturing capacity around the world.

But the US government is trying to change that. Through high tariffs on imports and hefty domestic tax credits, it is trying to make the cost of manufacturing solar panels in the US competitive enough for companies to want to come back and set up factories. The International Energy Agency has forecast that by 2027, solar-generated energy will be the largest source of power capacity in the world, exceeding both natural gas and coal—making it a market that already attracts over $300 billion in investment every year.

To understand the chances that the US will succeed, MIT Technology Review spoke to Shawn Qu. As the founder and chairman of Canadian Solar, one of the largest and longest-standing solar manufacturing companies in the world, Qu has observed cycle after cycle of changing demand for solar panels over the last 28 years. 

After decades of mostly manufacturing in Asia, Canadian Solar is pivoting back to the US because it sees a real chance for a solar industry revival, mostly thanks to the Inflation Reduction Act (IRA) passed in 2022. The incentives provided in the bill are just enough to offset the higher manufacturing costs in the US, Qu says. He believes that US solar manufacturing capacity could grow significantly in two to three years, if the industrial policy turns out to be stable enough to keep bringing companies in. 

How tariffs forced manufacturing capacity to move out of China

There are a few important steps to making a solar panel. First silicon is purified; then the resulting polysilicon is shaped and sliced into wafers. Wafers are treated with techniques like etching and coating to become solar cells, and eventually those cells are connected and assembled into solar modules.

For the past decade, China has dominated almost all of these steps, for a few reasons: low labor costs, ample supply of proficient workers, and easy access to the necessary raw materials. All these factors make made-in-China solar modules extremely price-competitive. By the end of 2024, a US-made solar panel will still cost almost three times as much as one produced in China, according to researchers at BloombergNEF. 

The question for the US, then, is how to compete. One tool the government has used since 2012 is tariffs. If a solar module containing cells made in China is imported to the US, it’s subject to as much as a 250% tariff. To avoid those tariffs, many companies, including Canadian Solar, have moved solar cell manufacturing and the downstream supply chain to Southeast Asia. Labor costs and the availability of labor forces are “the number one reason” for that move, Qu says.

When Canadian Solar was founded in 2001, it made all its solar products in China. By early 2023, the company had factories in four countries: China, Thailand, Vietnam, and Canada. (Qu says it used to manufacture in Brazil and Taiwan too, but later scaled back production in response to contracting local demand.)

But that equilibrium is changing again as further tariffs imposed by the US government aim to force supply chains to move out of China. Starting in June 2024, companies importing silicon wafers from China to make cells outside the country will also be subject to tariffs. The most likely solution for solar companies would be to “set up wafer capacity or set up partnerships with wafer makers in Southeast Asia,” says Jenny Chase, the lead solar analyst at BloombergNEF.

Qu says he’s confident the company will meet the new requirements for tariff exemption after June. “They gave the industry about two years to adapt, so I believe most of the companies, at least the tier-one companies, will be able to adapt,” he says.

The IRA, and moving the factories to the US

While US policies have succeeded in turning Southeast Asia into a solar manufacturing hot spot, not much of the supply chain has actually come back to the US. But that’s slowly changing thanks to the IRA, introduced in 2022. The law will hand out tax credits for companies producing solar modules in the US, as well as those installing the panels. 

The credits, Qu says, are enough to make Canadian Solar move some production from Southeast Asia to the US. “According to our modeling, the incentives provided just offset the cost differences—labor and supply chain—between Southeast Asia and the US,” he says.

Jesse Jenkins, an assistant professor in energy and engineering at Princeton University, has come to the same conclusion through his research. He says that the IRA subsidies and tax credits should offset higher costs of manufacturing in the US. “That should drive a significant increase in demand for made-In-America solar modules and subcomponents,” Jenkins says. And the early signs point that way too: since the introduction of the IRA, solar companies have announced plans to build over 40 factories in the US.

In 2023, Canadian Solar announced it would build its first solar module plant in Mesquite, Texas, and a solar cell plant in Jeffersonville, Indiana. The Texas factory started operating in late 2023, while the Indiana one is still in the works. 

The remaining challenges

While the IRA has brought new hope to American solar manufacturing, there are still a few obstacles ahead.

Qu says one big challenge to getting his Texas factory up and running is the lack of experienced workers. “Let’s face the reality: there was almost no silicon-based solar manufacturing in the US, so it takes time to train people,” he says. That’s a process that he expects to take at least six months. 

Another challenge to reshoring solar manufacturing is the uncertainty about whether the US will keep heavily subsidizing the clean energy industry, especially if the White House changes hands after the election this year. “The key is stability,” Qu says, “Sometimes politicians are swayed by special-interest groups.”

“Obviously, if you build a factory, then you do want to know that the incentives to support that factory will be there for a while,” says Chase. There are some indications that support for the IRA won’t necessarily be swayed by the elections. For example, jobs created in the solar industry would be concentrated in red states, so even a Republican administration would be motivated to maintain them. But there’s no guarantee that US policies won’t change course.

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

I don’t know about you, but I only learned last week that there’s something connecting MSG and computer chips.

Inside most laptop and data center chips today, there’s a tiny component called ABF. It’s a thin insulating layer around the wires that conduct electricity. And over 90% of the materials around the world used to make this insulator are produced by a single Japanese company named Ajinomoto, more commonly known for commercializing the seasoning powder MSG in 1909.

Hold on, what? 

As my colleague James O’Donnell explained in his story last week, it turns out Ajinomoto figured out in the 1990s that a chemical by-product of MSG production can be used to make insulator films, which proved to be essential for high-performance chips. And in the 30 years since, the company has totally dominated ABF supply. The product—Ajinomoto Build-up Film—is even named after it.

James talked to Thintronics, a California-based company that’s developing a new insulating material it hopes could challenge Ajinomoto’s monopoly. It already has a lab product with impressive attributes but still needs to test it in manufacturing reality.

Beyond Thintronics, the struggle to break up Ajinomoto’s monopoly is not just a US effort.

Within China, at least three companies are also developing similar insulator products. Xi’an Tianhe Defense Technology, which makes products for both military and civilian use, introduced its take on the material, which it calls QBF, in 2023; Zhejiang Wazam New Material and Guangdong Hinno-tech have also announced similar products in recent years. But all of them are still going through industrial testing with chipmakers, and few have recent updates on how well these materials have performed in mass-production settings.

“It’s interesting that there’s this parallel competition going on,” James told me when we recently discussed his story. “In some ways, it’s about the materials. But in other ways, it’s totally shaped by government funding and incentives.”

For decades, the fact that the semiconductor supply chain was in a few companies’ hands was seen as a strength, not a problem, so governments were not concerned that one Japanese company controlled almost the entire supply of ABF. Similar monopolies exist for many other materials and components that go into a chip.

But in the last few years, both the US and Chinese governments have changed that way of thinking. And new policies subsidizing domestic chip manufacturing are creating a favorable environment for companies to challenge monopolies like Ajinomoto’s.

In the US, this trend is driven by the fear of supply chain disruptions and a will to rebuild domestic semiconductor manufacturing capabilities. The CHIPS Act was announced to inject investment into chip companies that bring their plants back to the US, but smaller companies like Thintronics could also benefit, both directly through funding and indirectly through the establishment of a US-based supply chain.

Meanwhile, China is being cornered by a US-led blockade to deny it access to the most advanced chip technologies. While materials like ABF are not restricted in any way today, the fact that one foreign company controls almost the entire supply of an indispensable material raises the stakes enough to make the government worry. It needs to find a domestic alternative in case ABF becomes subject to sanctions too.

But it takes a lot more than government policies to change the status quo. Even if these companies are able to find alternative materials that perform better than ABF, there’s still an uphill battle to convince the industry to adopt it en masse.

“You can look at any dielectric film supplier (many from Japan and some from the US), and they have all at one time or another tried to break into ABF market dominance and had limited success,” Venky Sundaram, a semiconductor researcher and entrepreneur, told James. 

It’s not as simple as just swapping out ABF and swapping in a new insulator material. Chipmaking is a deeply intricate process, with components closely depending on each other. Changing one material could require a lot more knock-on changes to other components and the entire process. “Convincing someone to do that depends on what relationships you have with the industry. These big manufacturing players are a little bit less likely to take on a small materials company, because any time they’re taking on new material, they’re slowing down their production,” James said.

As a result, Ajinomoto’s market monopoly will probably remain while other companies keep trying to develop a new material that significantly improves on ABF. 

That result, however, will have different implications for the US and China. 

The US and Japan have long had a strategic technological alliance, and that could be set to deepen because both of them consider the rise of China a threat. In fact, Japan’s prime minister, Fumio Kishida, was just visiting the US last week, hoping to score more collaborations on next-generation chips. Even though there has been some pushback from the Japanese chip industry about how strict US export restrictions could become, this hasn’t been strong enough to sway Japan to China’s side.

All these factors give the Chinese government an even greater sense of urgency to become self-sufficient. The country has already been investing vast sums of money to that end, but progress has been limited, with many industry insiders pessimistic about whether China can catch up fast enough. If Ajinomoto’s failed competitors in the past tell us anything, it’s that this will not be an easy journey for China either.

Do you think China has a chance of cracking Ajinomoto’s monopoly over this very specific insulating material? Let me know your thoughts at zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. Following the explosive popularity of minute-long short dramas made for phones, China’s culture regulator will soon announce new regulations that tighten its control of them. (Sixth Tone)

• This is not a surprise to the companies involved. Some Chinese short-drama companies have already started to expand overseas, driven out by domestic policy pressures. I profiled one named FlexTV. (MIT Technology Review)

2. There have been many minor conflicts between China and the Philippines recently over maritime territory claims. Here’s what it feels like to live on one of those contested islands. (NPR)

3. The Chinese government has asked domestic telecom companies to replace all foreign chips by 2027. It’s a move that mirrors previous requests from the US to replace all Huawei and ZTE equipment in telecom networks. (Wall Street Journal $)

4. A decade ago, about 25,000 American students were studying in China. Today, there are only about 750. It may be unsurprising given recent geopolitical tensions, but neither country is happy with the situation. (Associated Press)

5. Latin America is importing large amounts of Chinese green technologies—mostly electric vehicles, lithium-ion batteries, and solar panels. (The Economist $)

6. China’s top spy agency says foreign agents have been trying to intercept information about the country’s rare earth industry. (South China Morning Post $)

7. Amid the current semiconductor boom, Southeast Asian youths are flocking to Taiwan to train and work in the chip industry. (Rest of World)

Lost in translation

The bodies of eight Chinese migrants were recently discovered on a beach in Mexico. According to Initium Media, a Singapore-based publication, this was the first confirmed shipwreck incident with Chinese migrants heading to the US, but many more have taken the perilous route in recent years. In 2023, over 37,000 Chinese people illegally entered the US through the border with Mexico.

The traffickers often arrange shabby boats with no safety measures to sail from Tapachula to Oaxaca, a popular route that circumvents police checkpoints on land but makes for an extremely dangerous journey often rocked by strong winds and waves. There had always been rumors of people going missing in the ocean, but these proved impossible to confirm, as no bodies were found. The latest tragedy was the first one to come to public attention. Of the nine Chinese migrants onboard the boat, only one survived. Three bodies remain unidentified today.

One more thing

Forget about the New York Times’ election-result needles and CNN’s relentless coverage by John King. In South Korea, the results of national elections are broadcast on TV with wild and whimsical animations. To illustrate the results of parliamentary elections that just concluded last week, candidates were shown fighting on a fictional train heading toward the National Assembly, parodying Mission: Impossible’s fight scene. According to the BBC, these election-night animations took a team of 70 to prepare in advance and about 200 people working on election night.

The best part of South Korean election night: the graphics. pic.twitter.com/XfFGkSD8k4

— Michelle Ye Hee Lee (@myhlee) April 10, 2024

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

If you’re a longtime subscriber to this newsletter, you know that I talk about China’s tech policies all the time. To me, it’s always a challenge to understand and explain the government’s decisions to bolster or suppress a certain technology. Why does it favor this sector instead of that one? What triggers officials to suddenly initiate a crackdown? The answers are never easy to come by.

So I was inspired after talking to Angela Huyue Zhang, a law professor in Hong Kong who’s coming to teach at the University of Southern California this fall, about her new book on interpreting the logic and patterns behind China’s tech regulations.

We talked about how the Chinese government almost always swings back and forth between regulating tech too much and not enough, how local governments have gone to great lengths to protect local tech companies, and why AI companies in China are receiving more government goodwill than other sectors today.

To learn more about Zhang’s fascinating interpretation of the tech regulations in China, read my story published today.

In this newsletter, I want to show you a particularly interesting part of the conversation we had, where Zhang expanded on how market overreactions to Chinese tech policies have become an integral part of the tech regulator’s toolbox today.

The capital markets, perpetually betting on whether tech companies are going to fare better or worse, are always looking for policy signals on whether China is going to start a new crackdown on certain technologies. As a result, they often overreact to every move by the Chinese government.

Zhang: “Investors are already very nervous. They see any sort of regulatory signal very negatively, which is what happened last December when a gaming regulator sent out a draft proposal to regulate and curb gaming activities. It just spooked the market. I mean, actually, that draft law is nothing particularly unusual. It’s quite similar to the previous draft circulated among the lawyers, and there are just a couple of provisions that need a little bit of clarity. But investors were just so panicked.”

That specific example saw nearly $80 billion wiped from the market value of China’s two top gaming companies. The drastic reaction actually forced China’s tech regulators to temporarily shelve the draft law to quell market pessimism. 

Zhang: If you look at previous crackdowns, the biggest [damage] that these firms receive is not in the form of a monetary fine. It is in the form of the [changing] market sentiment. 

What the agency did at that time was deliberately inflict reputational damage on [Alibaba] by making this surprise announcement on its website, even though it was just one sentence saying “We are investigating Alibaba for monopolistic practice.” But they already caused the market to panic. As soon as they made the announcement, it wiped off $100 billion market cap from this firm overnight. Compared with that, the ultimate fine of $2.8 billion [that Alibaba had to pay] is nothing.

China’s tech regulators use the fact that the stock market predictably overreacts to policy signals to discipline unruly tech companies with minimum effort.

Zhang: These agencies are very adept at inflicting reputational damage. That’s why the market sentiment is something that they like to [utilize], and that kind of thing tends to be ignored because people tend to fix any attention on the law.

But playing the market this way is risky. As in the previously mentioned example of the video-game policy, regulators can’t always control how significant the overreactions become, so they risk inflicting broader economic damage that they don’t want to be responsible for.

Zhang: They definitely learned how badly investors can react to their regulatory actions. And if anything, they are very cautious and nervous as well. I think they will be risk-averse in introducing harsh regulations.

I also think the economic downturn has dampened the voices of certain agencies that used to be very aggressive during the crackdown, like the Cyberspace Administration of China. Because it seems like what they did caused tremendous trauma for the Chinese economy.

The fear of causing negative economic fallout by introducing harsh regulatory measures means these government agencies may turn to softer approaches, Zhang says. 

Zhang: Now, if they want to take a softer approach, they would have a cup of tea with these firms and say “Here’s what you can do.” So it’s a more consensual approach now than those surprise attacks.

Do you agree that Chinese regulators have learned to take a softer approach to disciplining tech companies? Let me know your thoughts at zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. Covert Chinese accounts are pretending to be Trump supporters on social media and stoking domestic divisions ahead of November’s US election, taking a page out of the Russian playbook in 2016. (New York Times $)

2. Tesla canceled long-promised plans to release an inexpensive car. Its Chinese rivals are selling EV alternatives at less than one-third the price of the cheapest Teslas. (Reuters $)

3. Donghua Jinlong, a factory in China that makes a nutritional additive called “high-quality industrial-grade glycine,” has unexpectedly become a meme adored by TikTok users. No one really knows why. (You May Also Like)

4. Joe Tsai, Alibaba’s chairman, said in a recent interview that he believes Chinese AI firms lag behind US peers “by two years.” (South China Morning Post $)

5. At first glance, a hacker behind a multi-year attempt to hack supply chains seemed to come from China. But details about the hacker’s work hours suggest that countries in Eastern Europe or the Middle East could be the real culprit. (Wired $)

6. While visiting China, US Treasury Secretary Janet Yellen said that she would not rule out potential tariffs on China’s green energy exports, including products like solar panels and electric vehicles. (CNBC)

Lost in translation

Hong Kong’s food delivery scene used to be split between the German-owned platform Foodpanda and the UK-owned Deliveroo. But the Chinese giant Meituan has been working since May 2023 on cracking into the scene with its new app KeeTa, according to the Chinese publication Zhengu Lab. It has so far managed to capture over 20% of the market.

Both of Meituan’s rivals waive the delivery fee only for larger orders, which makes it hard for people to order food alone. So Meituan decided to position itself as the platform for solo diners by waiving delivery fees for most restaurants, saving users up to 30% in costs. To compete with the established players, the company also pays higher wages to delivery workers and charges lower commission fees to restaurants. 

Compared with mainland China, Hong Kong has a tiny delivery market. But Meituan’s efforts here represent a first step as it works to expand into more countries overseas, the company has said.

One more thing

For some hard-to-explain reasons, every time US Treasury Secretary Janet Yellen visits China, Chinese social media becomes obsessed with what and how she eats during the trip. This week, people were zooming into a seven-second video of Yellen’s dinner to scrutinize … her chopstick skills. Whyyyyyyy?

The way China regulates its tech industry can seem highly unpredictable. The government can celebrate the achievements of Chinese tech companies one day and then turn against them the next.

But there are patterns in how China approaches regulating tech, argues Angela Huyue Zhang, a law professor at Hong Kong University and author of the new book High Wire: How China Regulates Big Tech and Governs Its Economy. The way Chinese policies change almost always follow a three-phase progression: a lax approach where companies are given relative flexibility to expand and compete, sudden harsh crackdowns that slash profits, and eventually a new loosening of restrictions. 

Take Alibaba and Tencent as examples. Since the 2000s, the two tech giants have made hundreds of mergers and investments, as a result of which their business empires expanded to include almost every aspect of digital life in China. This insatiable expansion came at the expense of users, who faced higher prices and less choice, but Chinese regulators let it slide. Then, suddenly, the government started a tech crackdown in 2020. All of a sudden, past mergers and acquisitions were under investigation, and hefty fines were meted out to punish the companies for antitrust violations, including a $2.8 billion fine for Alibaba. 

MIT Technology Review recently spoke with Zhang about her new book and how to apply her insights to China’s tech industry, including significant new sectors like artificial intelligence.

The pendulum swing

“There’s this saying I also cited in my book: 一放就乱, 一抓就死 (loosening causes chaos; tightening up causes death),” Zhang says. The Chinese expression perfectly captures how the regulators dramatically yet predictably oscillate between doing too little to police the tech sector and doing too much. 

In the book, Zhang argues that Chinese tech platforms have long been accused of obstructing competition, infringing on privacy, and violating the labor rights of gig workers—but regulators accommodated them in all three areas until suddenly putting the companies under scrutiny in late 2020. And after the peak of enforcement in 2022, the regulators slowed down on all three fronts and reached a compromise with Chinese companies. 

Outside the examples in the book, “I think [the pattern] fits almost every sector,” Zhang says. From financial innovations like peer-to-peer loans in the mid-2010s to online tutoring, which exploded in popularity during the pandemic, they all went through similar shifts in experience with the regulators.

The government can be a helping hand

Western observers of Chinese policies often focus on the crackdown phase. Historically, it’s involved some dramatic moments—for example, the government forcing the ride-hail giant Didi to delist from the New York Stock Exchange or slapping antitrust fines on Alibaba after its former head, Jack Ma, made a public speech against regulation. 

But Zhang warns that these high-profile crackdowns mask the symbiotic relationship between tech companies and the government. “We tend to see [Chinese tech regulations] as very predatory,” she says, but “regulations actually give a helping hand to these firms.”

For many government officials, especially at the provincial and local levels, tech companies are the most important contributors to tax revenues and employment. They are often referred to as “local champions” or “little giants,” and their business interests are directly tied to the interests of local governments. In turn, the governments often go to great lengths to protect these companies. 

Zhang found, for example, that Chinese local courts spend tremendous judicial resources helping tech firms resolve online disputes. In a six-year span from 2016 to 2021, three local courts in the southern city of Guangzhou processed over 130,000 cases where the Chinese fintech company Lakala sued its users for defaulting on microloans. At one of the courts, each judge on average ruled on about 1,400 Lakala disputes in 2019 or 2020, meaning the court was essentially turned into an outsourced dispute arbitrator for the company. And the result? Lakala won almost all the cases.

When interests align on AI

Currently, AI is making the case that the interests of the government and the Chinese companies are aligned even more closely. 

That’s because the technology is seen as crucial to achieving China’s goals of technological supremacy and self-sufficiency, Zhang says. 

During China’s post-2008 economic slump, consumer tech startups and the platform economy were seen as a way to inspire new growth; the same is happening with AI today. At China’s annual parliamentary meeting last month, President Xi Jinping coined the term “new quality productive forces,” meaning the new sectors that are expected to counter China’s current economic slowdown. And a campaign focused on AI was explicitly mentioned in this connection.

“It’s a business that the Chinese government is deeply involved in from the start,” Zhang says. The Chinese government has taken multiple roles in the development of AI, functioning as policy maker, incubator, investor in AI startups, supplier of research, customer of AI applications, and more. “And now behind every successful Chinese AI firm, there is a local government,” she says. “That powerful backing will offer more political protection for Chinese AI businesses.”

The AI honeymoon phase today

The government’s deeply embedded interest in China’s AI industry means that the industry will stay in that initial phase of lax regulation for a while, Zhang says. And she argues that AI regulations in China today are looser than those in the US and Europe.

This claim may seem a little counterintuitive at first. While the EU has indeed led the world in passing AI regulation, China has reacted much more swiftly than the US, including passing some sweeping regulations about generative AI, deepfakes, and recommendation algorithms in the past two years.

But Zhang believes that these regulations are strict only when it comes to freedom of speech and content control, areas in which the Chinese government has been increasingly stringent. Other than that, the recent regulations offer vague principles and few enforceable measures to prevent the AI from causing harm, including harm to Chinese people’s human rights.

The Cyberspace Administration of China (CAC), an internet regulator that has a close relationship with the Communist Party’s propaganda bureau, has been in charge of reducing the risk that generative AI models will produce politically damaging content. Some of its restrictive measures, like requiring language models to reflect socialist values and asking for real-identity verification for users, will no doubt make it harder for Chinese companies to innovate and compete. But the CAC’s work often clashes with the priorities of other Chinese government agencies like the Ministry of Industry and Information Technology, which are more focused on boosting China’s technology capabilities.

Judging from Chinese AI regulations so far, the pro-growth faction has prevailed, says Zhang. “At least you [in the US] have the FTC open an investigation into OpenAI. In China, did you see the CAC open an investigation into Baidu or ByteDance? No. And I predict they are very unlikely to do that in the future, unless something really bad happens,” she says.

How bad would it have to be to trigger the switch to regulatory crackdown? Zhang says it would take a big AI misuse that sets off wide-ranging controversies and threatens social stability. If that happens, then the Chinese regulatory pendulum will dutifully swing to the harsh side again.

When it happens, it will be quick. “It will be quite random and quite sudden,” Zhang says, “and it will be a surprise.”

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

Like most reporters, I have accounts on every social media platform you can think of. But for the longest time, I was not on Threads, the rival to X (formerly Twitter) released by Meta last year. The way it has to be tied to your Instagram account didn’t sit well with me, and as its popularity dwindled, I felt maybe it was not necessary to use it.

But I finally joined Threads last week after I discovered that the app has unexpectedly blown up among Taiwanese users. For months, Threads has been the most downloaded app in Taiwan, as users flock to the platform to talk about politics and more. I talked to academics and Taiwanese Threads users about why the Meta-owned platform got a redemption arc in Taiwan this year. You can read what I discovered here.

I first noticed the trend on Instagram, which occasionally shows you a few trending Threads posts to try to entice you to join. After seeing them a few times, I realized there was a pattern: most of these were written by Taiwanese people talking about Taiwan.

That was a rare experience for me, since I come from China and write primarily about China. Social media algorithms have always shown me accounts similar to mine. Although people from mainland China, Hong Kong, and Taiwan all write in Chinese, the characters we use and the expressions we choose are quite different, making it easy to spot your own people. And on most platforms that are truly global, the conversations in Chinese are mostly dominated by people in or from mainland China, since its population far outnumbers the rest. 

As I dug into the phenomenon, it soon turned out that Threads’ popularity has been surging at an unparalleled pace in Taiwan. Adam Mosseri, the head of Instagram, publicly acknowledged that Threads has been doing “exceptionally well in Taiwan, of all places.” Data from Sensor Tower, a market intelligence firm, shows that Threads has been the most downloaded social network app on iPhone and Android in Taiwan almost every single day of 2024. On the platform itself, Taiwanese users are also belatedly realizing their influence when they see that comments under popular accounts, like a K-pop group, come mostly from fellow Taiwanese users. 

But why did Threads succeed in Taiwan when it has failed in so many other places? My interviews with users and scholars revealed a few reasons.

First, Taiwanese people never really adopted Twitter. Only 1% to 5% of them regularly use the platform, now called X, estimates Austin Wang, a political science professor at the University of Nevada, Las Vegas. The mainstream population uses Facebook and Instagram, but still yearns for a platform for short text posts. The global launch of Threads basically gave these users a good reason to try out a Twitter-like product.

But more important, Taiwan’s presidential election earlier this year means there was a lot to talk, debate, and commiserate about. Starting in November, many supporters of Taiwan’s Democratic Progressive Party (DPP) “gathered to Threads and used it as a mobilization tool,” Wang says. “Even DPP presidential candidate Lai received more interaction on Threads than Instagram and Facebook.” 

It turns out that even though Meta has tried to position Threads as a less political version of X, what actually underpinned its success in Taiwan was still the universal desire to talk about politics.

“Taiwanese people gather on Threads because of the freedom to talk about politics [here],” Liu, a designer in Taipei who joined in January because of the elections, tells me. “For Threads to depoliticize would be shooting itself in the foot.” 

The fact that there are an exceptionally large number of Taiwanese users on Threads also makes it a better place to talk about internal politics, she says, because it won’t easily be overshadowed or hijacked by people outside Taiwan. The more established platforms like Facebook and X are rife with bots, disinformation campaigns, and controversial content moderation policies. On Threads there’s minimal interference with what the Taiwanese users are saying. That feels like a fresh breeze to Liu.

But I can’t help feeling that Threads’ popularity in Taiwan could easily go awry. Meta’s decision to keep Threads distanced from political content is one factor that could derail Taiwanese users’ experience; an influx of non-Taiwanese users, if the platform actually manages to become more successful and popular in other parts of the world, could also introduce heated disagreements and all the additional reasons why other platforms have deteriorated. 

These are some tough questions to answer for Meta, because users will simply flow to the next trendy, experimental platform if Threads doesn’t feel right anymore. Its success in Taiwan so far is a rare win for the company, but preserving that success and replicating it elsewhere will require a lot more work.

Do you believe Threads stands a chance of rivaling X (Twitter) in places other than Taiwan? Let me know your thoughts at zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. Morris Chang, who founded the Taiwan Semiconductor Manufacturing Company at the age of 55, is an outlier in today’s tech industry, where startup founders usually start in their 20s. (Wall Street Journal $)

2. A group of Chinese researchers used the technology behind hypersonic missiles to make high-speed trains safer. (South China Morning Post $)

3. The US government is considering cutting the so-called de minimis exemption from import duties, which makes it cheap for Temu and Shein to send packages to the US. But lots of US companies also benefit from the exemption now. (The Information $)

4. The Chinese commerce minister will visit Europe soon to plead his country’s case amid the European Commission’s investigation into Chinese electric vehicles. (Reuters $)

5. After three years of unsuccessful competition with WhatsApp, ByteDance’s messaging app designed for the African market finally shut down last month. (Rest of World)

6. The rapid progress of AI makes it seem less necessary to learn a foreign language. But there are still things AI loses in translation. (The Atlantic $)

7. This is the incredible story of a Chinese man who takes his piano to play outdoors at places of public grief: in front of the covid quarantine barriers in Wuhan, at the epicenter of an earthquake, on a river that submerged villages. And he plays the same song—the only song he knows, composed by the Japanese composer Ryuichi Sakamoto. (NPR)

Lost in translation

With Netflix’s March release of The Three Body Problem, a series adapted from the global hit sci-fi novel by Chinese author Liu Cixin, Western audiences are also learning about a movie-like real-life drama behind the adaptation. In 2021, the Chinese publication Caixin first investigated the mysterious death of Lin Qi, a successful businessman who bought the movie rights to the book. In 2017, he hired Xu Yao, a prominent attorney, to work on legal affairs and government relations.

In December 2020, Lin died after he was poisoned by a mysterious mix of toxins. According to Caixin, Xu is a fan of the TV series Breaking Bad and had his own plant in Shanghai where he made poisons. He would order hundreds of different toxins through the dark web, mix them, and use them on pets to experiment. A week before Lin’s death, Xu gave him a bottle of pills that were supposedly prebiotics, but he had replaced them with poison. 

Xu was arrested soon after Lin died, and he was sentenced to death on March 22 this year.

One more thing

Taobao, China’s leading e-commerce platform, announced it’s experimenting with delivering packages by rockets. Yes, rockets. Made by a Chinese startup, Taobao’s pilot rockets will be able to deliver something as big as a car or a truck, and the rockets can be reused for the next delivery. To be honest, I still can’t believe this wasn’t an April Fool’s joke.

For most people around the world, Meta’s text-based social network Threads is a platform that they haven’t thought of for months. But for Liu, a design professional in Taipei, it’s where she’s receiving unprecedented attention. 

“My casual posts often receive a large number of reposts now. It used to only happen every few months on Twitter, but it’s happening every few weeks or even days on Threads,” says Liu, who has used Twitter (now renamed X) for more than eight years and has posted on Threads since January. She asked MIT Technology Review to use only her last name for privacy reasons.

She’s not the only person feeling this surge of popularity. While most users left Threads soon after its launch and meteoric rise in July 2023, in Taiwan people have recently started to come back to the platform. There, Threads has dominated app-store download charts for months. Prominent officials have set up accounts, and it’s become the most popular platform among young people.

Even Meta has noticed the pickup in interest. In early March, Adam Mosseri, the head of Instagram, shared in an “ask me anything”–style story that “[Threads] is doing really well in a variety of countries, exceptionally well in Taiwan of all places, which has been fun to see.” A Meta spokesperson confirms that Mosseri has publicly spoken about the trend but declined to offer more data on the platform’s growth in Taiwan.

Users and observers point to a few factors that contributed to Threads’ unexpected success on the island, including the fact that Twitter never became truly mainstream for Taiwanese people. Threads has managed to meet the demand for open discussion when Meta’s other platforms, like Facebook, are losing their appeal. Taiwan’s presidential election in January also brought in a significant number of new accounts and a lively discussion of politics and social issues.

As a result, many people in Taiwan are joining Threads and using it daily. Liu spends less than an hour on average every day on the app, where she writes down whatever’s on her mind. Originally, her friends were real-life acquaintances connected through Instagram, but she’s increasingly making new friends on the platform now. 

“I’m an ordinary, introverted person … I feel so surprised and honored for the high level of attention I receive [on Threads]. This has never happened on any other platform,” she says.

The elections gave Threads a second chance

Threads was introduced to the world as Meta’s answer to Twitter after the latter was infamously acquired by Elon Musk, prompting many long-term users to look for alternatives. But in Taiwan, unlike most other places that began experimenting with Threads, people had never really adopted Twitter in the first place. “According to numerous surveys, at most 1% to 5% of Taiwanese people use Twitter regularly,” Austin Wang, an assistant professor of political science at the University of Nevada, Las Vegas, said in an email. 

There were a few exceptions. “I use [Twitter] first because the K-pop circles use it to save images of their idols, and secondly because LGBT communities (especially gay men) use it as a subculture social platform to meet new people,” says Sebastian Huang, a college student in Taipei. 

Outside these niche groups, though, Threads had a fresh chance to win Taiwanese users over. “In my observation, [Threads] popularized Twitter’s socialization logic and pushed it toward the mainstream communities,” Huang says.

Still, Threads’ popularity plummeted after its launch in July 2023. In Taiwan—like the rest of the world—many users left the platform after satisfying their initial curiosity. 

But the 2024 Taiwanese presidential election gave it another chance. Wang, who studies social media in Taiwan, traced the platform’s second rise to November of last year, starting with the supporters of Taiwan’s Democratic Progressive Party (DPP), often associated with the color green. “Many (worried) pan-green supporters noticed that their complaints on politics were promoted to more readers on Threads than any other social media platforms (especially Facebook and Instagram), so more and more pan-green supporters gathered to Threads and used it as a mobilization tool,” he says.

The election concluded in mid-January, with DPP candidate Lai Ching-te elected as Taiwan’s president. Many supporters of his party stayed on the platform. And as it became influential, other political figures also reactivated their Threads accounts and started posting regularly, trying to join the conversation. Everyday users who are less interested in politics came along too.

On almost every day of the past three months, Threads has been the most downloaded social network app in both Apple’s and Android’s app stores in Taiwan, according to Sensor Tower, an app store intelligence firm. It surpassed both Western social platforms and those popular in China. 

What does Taiwan Threads look like?

Wang, who has been actively posting on Threads and accumulated over 3,000 followers, observes that there are two major demographics among Taiwan’s Threads users today: the pro-green voters, and younger students who are still in middle school and high school. “In recent weeks, there is a considerable amount of discussion on how to choose colleges, majors, and even high schools,” he says.

Since Threads doesn’t have an official name in Chinese, Taiwanese users have tried to translate it in creative ways. Some stay close to the meaning and call it 串 or chuan, which means a string of beads or other objects (it could also mean a kebab skewer). Others call it 脆 or cui, which means crispy or fragile. It’s a transliteration attempt that many feel is too far-fetched, but since there’s no sound like “th” in Mandarin, it’s the best alternative, and it has already caught on among the users and surpassed other names. 

What defines the content on Threads is a mix of political and lifestyle posts. On the one hand, some of the most influential accounts are Taiwanese politicians at all levels, including the presidential candidates. On the other, Threads users have embraced a type of content called 廢文—a cross between trash talk and light-stakes monologue. 

As a result, to gain a following on Threads, the best practice is to mix up the serious and the unserious. One local representative candidate became unexpectedly famous when people discovered that his son was physically attractive. Joking about how this son’s virality has eclipsed his own, the politician now calls himself “The father of the son of Phoenix Cheng” on Threads, where he has over 268,000 followers.

“People like Phoenix Cheng like to post 廢文—talking about private matters in a humorous way. It shows you an unusual side of them,” says Jung-Chin Shen, a professor of international business at Fu Jen Catholic University in New Taipei City. 

Taiwanese politicians typically put their serious policy messaging on Facebook, but it’s not wise for them to approach Threads in the same way. “If I have followed your Facebook accounts already, why would I want to read the same thing on another platform for a second time?” Shen says. “Official, serious debates and political messaging can’t appeal to Threads users anymore.”

A delicate balance on political content

The success of Threads in Taiwan shows that politics is still one of the main reasons people come to a text-based social network, but it also highlights Meta’s uncomfortable relationship with political content on its platforms.

Before the emergence of Threads, these discussions happened mostly on Facebook and Twitter, but the prevalence of bots, misinformation, disinformation, and spam content drove people to find new alternatives.

Liu, who joined Threads in January because of the election-related content, says talking about politically sensitive topics on other social platforms could often result in being shadowbanned or even suspended. Threads, with its minimal political moderation efforts so far, appeals to those looking for a place to discuss politics freely. 

“Taiwanese people gather on Threads because of the freedom to talk about politics [here],” she says.

In turn, these political discussions have made the platform popular, at least for now. “The presidential elections in Taiwan have high mobilization and receive a high level of discussion on social media,” says Shen. “Other than in times like this, it’s rare to have a lot of people migrate to a new platform in a short amount of time.”

For Threads, Taiwan presents both an opportunity and a challenge to its current content policies. The blending of politics and lighthearted content has been a successful example for the platform, which was pitched from the very beginning as a less political, less serious alternative to Twitter. But it may want to deemphasize politics even more. In February, the month after Taiwan’s election results, Meta confirmed its position that Instagram and Threads “won’t proactively recommend content about politics.” Such material will be hidden in some recommendation features by default, and the reach of users talking about politics will be severely restricted.

“As with all our products, we take safety seriously, and we enforce Instagram’s Community Guidelines on content and interactions in Threads,” a Meta spokesperson said in response to MIT Technology Review’s emailed questions. The company’s third-party fact-checking partners “will soon be able to review and rate false content that originates on Threads,” she says. The company didn’t answer questions about where it draws the line between political and nonpolitical content when it comes to content recommendation. 

Those who came for the political content are pretty pessimistic about the future of Threads if it carries out this change. “For Threads to depoliticize would be shooting itself in the foot,” Liu says.

Even users who are on Threads for different reasons don’t necessarily think the platform should take a blanket approach. Huang, the college student, says he’s not a political person and doesn’t want to talk about politics all the time. He registered his Threads account anonymously, intentionally separating it from his real-life acquaintances. In fact, he mutes anyone who talks about politics on his Threads timeline.

“But I also feel like it’s not the best solution to straight-up restrict [political content],” he says. “It’s better if users can control it by themselves.”

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

In 2019, MIT Technology Review covered a report that shined a light on how fast China’s AI talent pool was growing. Its main finding was pretty interesting: the number of elite AI scholars with Chinese origins had multiplied by 10 in the previous decade, but relatively few of them stayed in China for their work. The majority moved to the US. 

Now the think tank behind the report has published an updated analysis, showing how the makeup of global AI talent has changed since—during a critical period when the industry has shifted significantly and become the hottest technology sector. 

The team at MacroPolo, the think tank of the Paulson Institute, an organization that focuses on US-China relations, studied the national origin, educational background, and current work affiliation of top researchers who gave presentations and had papers accepted at NeurIPS, a top academic conference on AI. Their analysis of the 2019 conference resulted in the first iteration of the Global AI Talent Tracker. They’ve analyzed the December 2022 NeurIPS conference for an update three years later.

I recommend you read the original report, which has a very well-designed infographic that shows the talent flow across countries. But to save you some time, I also talked to the authors and highlighted what I think are the most surprising or important takeaways from the new report. Here are the four main things you need to know about the global AI talent landscape today. 

1.  China has become an even more important country for training AI talent.

Even in 2019, Chinese researchers were already a significant part of the global AI community, making up one-tenth of the most elite AI researchers. In 2022, they accounted for 26%, almost dethroning the US (American researchers accounted for 28%). 

“Timing matters,” says Ruihan Huang, senior research associate at MacroPolo and one of the lead authors. “The last three years have seen China dramatically expand AI programs across its university system—now there are some 2,000 AI majors—because it was also building an AI industry to absorb that talent.” 

As a result of these university and industry efforts, many more students in computer science or other STEM majors have joined the AI industry, making Chinese researchers the backbone of cutting-edge AI research.

2. AI researchers now tend to stay in the country where they receive their graduate degree. 

This is perhaps intuitive, but the numbers are still surprisingly high: 80% of AI researchers who went to a graduate school in the US stayed to work in the US, while 90% of their peers who went to a graduate school in China stayed in China.

In a world where major countries are competing with each other to take the lead in AI development, this finding suggests a trick they could use to expand their research capacity: invest in graduate-level institutions and attract overseas students to come. 

This is particularly important in the US-China context, where the souring of the relationship between the two countries has affected the academic field. According to news reports, quite a few Chinese graduate students have been interrogated at the US border or even denied entry in recent years, as a Trump-era policy persisted. Along with the border restrictions imposed during the pandemic years, this hostility could have prevented more Chinese AI experts from coming to the US to learn and work. 

3. The US still overwhelmingly attracts the most AI talent, but China is catching up.

In both 2019 and 2022, the United States topped the rankings in terms of where elite AI researchers work. But it’s also clear that the distance between the US and other countries, particularly China, has shortened. In 2019, almost three-fifths of top AI researchers worked in the US; only two-fifths worked here in 2022. 

“The thing about elite talent is that they generally want to work at the most cutting-edge and dynamic places. They want to do incredible work and be rewarded for it,” says AJ Cortese, a senior research associate at MacroPolo and another of the main authors. “So far, the United States still leads the way in having that AI ecosystem—from leading institutions to companies—that appeals to top talent.”

In 2022, 28% of the top AI researchers were working in China. This significant portion speaks to the growth of the domestic AI sector in China and the job opportunities it has created. Compared with 2019, three more Chinese universities and one company (Huawei) made it into the top tier of institutions that produce AI research. 

It’s true that most Chinese AI companies are still considered to lag behind their US peers—for example, China usually trails the US by a few months in releasing comparable generative AI models. However, it seems like they have started catching up.

4. Top-tier AI researchers now are more willing to work in their home countries.

This is perhaps the biggest and also most surprising change in the data, in my opinion. Like their Chinese peers, more Indian AI researchers ended up staying in their home country for work.

In fact, this seems to be a broader pattern across the board: it used to be that more than half of AI researchers worked in a country different from their home. Now, the balance has tipped in favor of working in their own countries. 

This is good news for countries trying to catch up with the US research lead in AI. “It goes without saying most countries would prefer ‘brain gain’ over ‘brain drain’—especially when it comes to a highly complex and technical discipline like AI,” Cortese says. 

It’s not easy to create an environment and culture that not only retains its own talents but manages to pull scholars from other countries, but lots of countries are now working on it. I can only begin to imagine what the report might look like in a few years.  

Did anything else stand out to you in the report? Let me know your thoughts by writing to zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. The Dutch prime minister will visit China this week to discuss with Chinese president Xi Jinping whether the Dutch chipmaking equipment company ASML can keep servicing Chinese clients. (Reuters $)

• Here’s an inside look into ASML’s factory and how it managed to dominate advanced chipmaking. (MIT Technology Review)

2. Hong Kong passed a tough national security law that makes it more dangerous to protest Beijing’s rule. (BBC)

3. A new bill in France suggests imposing hefty fines on Shein and similar ultrafast-fashion companies for their negative environmental impact—as much as $11 per item that they sell in France. (Nikkei Asia)

4. Huawei filed a patent to make more advanced chips with a low-tech workaround. (Bloomberg $)

• Meanwhile, a US official accused the Chinese chip foundry SMIC of breaking US law by making a chip for Huawei. (South China Morning Post $)

5. Instead of the usual six and a half days a week, Tesla has instructed its Shanghai factory to reduce production to five days a week. The slowdown of EV sales in China could be the reason. (Bloomberg $)

6. TikTok is still having plenty of troubles. A new political TV ad (paid for by a mysterious new nonprofit), playing in three US swing states, attacks Zhang Fuping, a ByteDance vice president that very few people have heard of. (Punchbowl News)

• As TikTok still hasn’t reached a licensing deal with Universal Music Group, users have had to get creative to find alternative soundtracks for their videos. (Billboard)

7. China launched a communications satellite that will help relay signals for missions to explore the dark side of the moon. (Reuters $)

Lost in translation

The most-hyped generative AI app in China these days is Kimi, according to the Chinese publication Sina Tech. Released by Moonshot AI, a Chinese “unicorn” startup, Kimi made headlines last week when it announced it had started supporting inputting text using over 2 million Chinese characters. (For comparison, OpenAI’s GPT-4 Turbo currently supports inputting 100,000 Chinese characters, while Claude3-200K supports about 160,000 characters.)

While some of the app’s virality can be credited to a marketing push that intensified recently. Chinese users are now busy feeding popular and classic books to the model and testing how well it can understand the context. Feeling threatened, other Chinese AI apps owned by tech giants like Baidu and Alibaba have followed suit, announcing that they will soon support 5 million or even 10 million Chinese characters. But processing large amounts of text, while impressive, is very costly in the generative AI age—and some observers worry this isn’t the commercial direction that companies ought to head in.

One more thing

Fluffy pajamas, sweatpants, outdated attire: young Chinese people are dressing themselves in “gross outfits” to work—an intentional provocation to their bosses and an expression of silent resistance to the trend that glorifies career hustle. “I just don’t think it’s worth spending money to dress up for work, since I’m just sitting there,” one of them told the New York Times.

Update: The story has been updated to clarify the affiliation of the report authors.

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

Over the last year, a few Chinese influencers have made millions of dollars peddling short video lessons on AI, profiting off people’s fears about the as-yet-unclear impact of the new technology on their livelihoods. 

But the platforms they thrived on have started to turn against them. Just a few weeks ago, WeChat and Douyin began suspending, removing, or restricting their accounts. While influencers on these platforms have been turning people’s anxiety into traffic and profits for a long time, the latest actions show how Chinese social platforms are trying to contain the damage before it goes too far. 

The backlash started last month, as students angrily complained on social media about the superficiality of the courses, saying that they fell far short of the educational promises made about them. 

“I paid 198 RMB ($27.50), and the first three courses were void of actual content. It’s all about urging people to keep paying 1980 RMB for the next course,” Bessie, a Chinese user of the social media site Xiaohongshu, posted about her experience. The courses were created by Li Yizhou, a serial entrepreneur turned startup mentor who, despite having no background in AI, pivoted to posting about explaining AI and drumming up anxiety after the release of ChatGPT in November 2022.

Li sold his entry-level course package for $27.50, and an advanced one for 10 times that price. The cheaper offering contained 40 lesson videos, most of which were around 10 minutes long. Li’s course consisted of tutorials of specific generative AI tools, talks with Chinese AI company executives, and introductions to unrelated topics like how to manage your time more effectively. 

His lessons were a huge commercial success. According to the social media data analysis site Feigua, they were sold over 250,000 times last year, which could have brought in over $6 million in revenue. 

Li is not the only influencer who, despite having no background in AI, saw a business opportunity to calm people’s AI anxieties with quick fixes. There’s also “Teacher He,” an influencer with over 7 million followers who until recently mostly talked about marketing and personal finance, and Zhang Shitong, also followed by millions, whose usual videos mix basic economics with sensational conspiracies like 9/11 denialism. These creators also offered beginner AI lessons at a similar price to Li’s.

In addition to quality complaints, buyers reported that it was hard to get a refund when they changed their mind. Bessie tells MIT Technology Review that she got a refund since she applied early, but others who applied for a refund more than a week after the purchase were denied. A Beijing-based AI community website has also accused Li of appropriating their free user-contributed templates and selling them for profit as part of his course offering. 

By late February, the platforms that hosted these video lessons began to heed the complaints. All of the classes by Li and other AI gurus have been removed from Chinese social media and e-commerce websites. Li hasn’t posted on any of his social media channels since he was suspended in late February. Other creators like “Teacher He” and Zhang Shitong have also been silent.

Li and “Teacher He” didn’t respond to a media inquiry sent by MIT Technology Review. But a customer representative working for Zhang Shitong said the team processes all refund requests in 12 hours and that it was the team’s own decision to not post anything for the past three weeks.

On Douyin, the Chinese version of TikTok, Li’s account, which used to have over 3 million followers, is now hidden from search results. WeChat Channels, another popular short-video platform, blocked Li and other similar creators from getting new followers in the last week of February. Other smaller platforms have also taken action. Zhishi Xingqiu, a Patreon-like platform that was used by many influencers to sell access to AI-focused communities, has now blocked the search for keywords like “AI,” “Li Yizhou,” or “Sora.”

But none of the platforms have specified which rules the gurus violated. While they may have overpromised with their marketing, it’s hard to say whether their activities really qualified as “scams.” Douyin and WeChat declined to comment on their decisions.

However, there are signs that the restrictions could be reversed. While Chinese social media platforms often permanently delete the accounts of users they believe are flouting rules, these AI course creators have kept their accounts on all platforms. On WeChat, after around two weeks of being blocked from receiving new followers, the creators quietly regained that ability in mid-March. On Douyin, Li’s account was hidden from in-app search results, but his past videos can still be found by going directly to his profile page. 

So far, the Chinese government has not directly addressed the phenomenon or given its official stance. The government has been reining in the livestreaming industry heavily in recent years to censor how influencers act and post, and Chinese platforms set their own rules accordingly, sometimes ahead of government orders, to show they are doing their parts in content regulation.

Even as the creators and their lessons were removed online, there are still lots of Chinese people keen to access these lessons. On social media, some people are now reselling pirated videos of Li’s courses through file sharing, likely without Li’s permission. Now, instead of $27.50, people can spend a few bucks to access the whole course package.

Do you think these AI gurus have crossed a line? Let me know your thoughts at zeyi@technologyreview.com.

---

Now read the rest of China Report

Catch up with China

1. The US House of Representatives voted overwhelmingly to pass a bill that would force ByteDance to either sell TikTok or see it banned in the US. Now it’s heading to the Senate, where there’s less urgency to pass it. (Associated Press)

2. While the TSMC chip plant in Arizona is delayed, the company’s other new plant in Japan is set to start mass production on schedule in the fourth quarter of 2024. (Wall Street Journal $)

3. Tesla is talking to countries like Thailand to prepare for a potential production expansion in Southeast Asia. But it will have to compete with Chinese EV companies like BYD, which currently accounts for over a quarter of the EVs sold in the region. (Reuters $)

4. An obscure Chinese e-commerce platform called Pandabuy is recruiting influencers to peddle counterfeit products on TikTok and Facebook. (Wired $)

5. The US and Chinese governments quietly renewed their bilateral deal on science and technology research for another six months. (Wall Street Journal $)

6. Chinese students and academics say they are increasingly being targeted at US airports when they enter the country. (Washington Post $)

7. As the Chinese population ages quickly, a tutoring industry for the elderly is thriving. (Reuters $)

Lost in translation

As the Chinese automobile industry moves fast toward battery-operated cars and electric motors, internal combustion engine technology is increasingly seen as a thing of the past. The Chinese publication Economic Observer talked to students who chose to study combustion engines out of their love for cars. They found it’s a decision some now regret, as they’re finding it hard to land a job after graduation. 

Engineering universities are recruiting experts who can teach students about car batteries, but the pace is not fast enough to catch up with the speed of the Chinese market. From January to July 2023, there was a 6% increase in job postings in the automotive industry in China, but a 18% increase in job postings in the EV industry. As a result, large numbers of combustion engineering students say they are being rejected by the auto industry. They either have to compete for the limited positions still available, or find jobs outside the car industry.

One more thing

Youdao, a Chinese online dictionary app, recently started letting users upload their own pronunciations of English words to appear alongside the standard pronunciations in American or British accents. It soon became a vehicle for fun, with people competing to insert jokes, cultural memes, viral TikTok soundbites, and dramatic acting as pronunciations. In a particularly amusing example, someone pronounces “constipation” as if they are actually experiencing it.

View this post on Instagram

A post shared by CANDISE LIN | CANTONESE TUTOR (@dr.candiselin)