When I talk to people about AI safety, I often hear this logic:
America must slow down AI development before we build a machine intelligence that escapes human control.
But we can’t slow down unless China slows down.
China won’t slow down.
Therefore we can’t slow down.
By this logic, the future of AI development—and, according to some, the future of the economy and the entire human race—depends on America’s understanding of China’s approach to artificial intelligence. But I cannot tell you how many times I have spoken to blazingly intelligent people around the AI industry who have told me “Now, I’m not a China expert, but…” and then gone on to defend America’s AI strategy in the context of a subject that they just told me they don’t fully understand.
I myself have never pretended to be a China expert. That’s why this article isn’t written by me, exactly. Instead, I spent the last two weeks calling and emailing a dozen experts on China and AI and reading the primary sources they sent me. Then I organized the most interesting parts of those interviews under seven big questions about China, AI, safety, and what the US government and the AI labs can and should do to protect the world. My summary analysis of the China-AI conundrum is at the bottom of this article.
Some context is useful here before we dive into the interviews. For a long time, several AI experts insisted that China wasn’t “AGI-pilled”—meaning that the nation’s government and labs weren’t devoting significant capital and compute toward the construction of world-leading models with the goal of producing a machine smarter than human experts across domains. But that certainty is weakening. A leaked transcript from a May 2026 closed-door investor meeting revealed that DeepSeek founder Liang Wenfeng is devoted to building AGI, and Chinese researchers are publishing more papers about what they call “L5” AI—a technology that could “govern” its own self-improvement without human intervention. Tempering these ambitions is the fact that China is significantly behind the US in compute and data center construction, and it’s unclear if the government sees AI as central to future geopolitical dominance as America does.
Matt Sheehan, senior fellow at the Carnegie Endowment for International Peace:
China and the U.S. have different stories about AI. Xi Jinping sees artificial intelligence as an across-the-board enabler of a huge number of industries. China’s biggest AI policy in recent years is called the “AI+” plan. It’s AI plus manufacturing, and AI plus scientific research. At this point, it is far and away the technology they care the most about. In the U.S., AGI has always been the anchor for how everybody sees AI. Every development is put into that narrative. Every safety incident and every big math problem solved immediately gets put into the context of “does this mean that AGI is close?” But China evaluates new pieces of evidence without the conclusion in mind. It’s not to say the safety people are wrong. But they’ve been drawn to that conclusion for a long time. China’s political leadership has not.
Jordan Schneider, founder and editor of China Talk:
This is sort of a hot take, but I think the people building AI in the U.S., and the West in general, have been deeply shaped by the science fiction stories that they read when they were younger about a post-AGI utopia. I’m thinking of stuff like The Player of Games, and writers like Iain M. Banks. These guys believe that they are building the future that they read about when they were kids. And that shapes the way they talk about the potential of the technology.
Martin Chorzempa, senior fellow at the Peterson Institute for International Economics:
I’m an outlier in thinking Chinese firms are more AGI/RSI-pilled than most people believe. DeepSeek says its goal is AGI. Moonshot says its goal is AGI and globalization. Other labs’ earnings calls talk about trying to unleash RSI. The Chinese labs see themselves as fast followers. From their perspective, the American labs are dealing with the hard challenges of alignment and security, and the Chinese labs think they can piggyback on that work.
In her essay “How China Hopes to Build AGI Through Self-Improvement,” Zilan Qian, a program associate at the Oxford China Policy Lab, writes that Chinese researchers tend to imagine AGI emerging slowly through robotics and via feedback from the physical world, whereas Americans are more likely to imagine a sudden and purely software-driven explosion of intelligence via recursive self-improvement. These different attitudes might be shaped by on-the-ground capabilities, such as the fact that China has fewer high-powered chips but has a deep robotics supply chain, whereas the U.S. has abundant compute but weaker on robotics. Here’s Qian:
Chinese scientists’ vision of AGI and self-improvement looks quite different from that of Silicon Valley. Rather than a rapid software-driven intelligence explosion — AI building AI in a recursive loop — Chinese thinking converges on something more embodied: human-level intelligence that requires physical-world interactions…
What matters is not only what Chinese AI researchers and Beijing believe AGI is, but also what happens quietly beneath those beliefs. Just as China sees the future of AI through its manufacturing strength and chip shortage, the U.S., with abundant chips and less manufacturing capabilities, sees a different version. The U.S. and China’s roads to AGI appear to be different, and perhaps the destinations do too. But if each side’s vision of AGI is shaped by what it already controls, then neither is well-positioned enough to recognize what the other is actually building.
Matt Sheehan:
The prisoner’s dilemma dynamic is real. At the same time, I have doubts about the claim that once AI passes some threshold of superintelligence, it will give China a “decisive strategic advantage,” which is the term floating around national-security AI circles. Treasury Secretary Scott Bessent has said if we lose AI to China, it’s all over, because once China achieves superintelligence, they can demolish our cyber and nuclear capabilities and put us in a permanent state of subservience. But when you talk to actual nuclear security people, they have a very hard time getting behind that idea. If you’re a country that invents a powerful tool that you don’t understand and don’t know if you can control, is the first thing you do going to be to declare war? I’m not sure about that.
Rush Doshi, former deputy senior director for China and Taiwan on Joe Biden’s National Security Council, currently at Georgetown University and the Council on Foreign Relations:
AI research [is] a bit like gain-of-function research in biology. That kind of research can be dangerous: if you increase the lethality or transmissibility of a virus that you’re studying, and it can be very damaging to everyone if it escapes containment. There’s an element of these models, particularly when it comes to cyber risk, that resembles gain-of-function risk, where the sandbox perhaps takes the place of the biolab. The problem in biology is that you cannot always get countries to stop doing gain-of-function research. So, yes, the world created some standards, but implementation remains a challenge. China, for example, has cut corners with safety in their biolabs.
Two weeks ago, the New York Times reported that China’s top spy chief warned the state that AI could pose a direct threat to the Chinese Communist Party’s hold on power and “pose a threat to humanity itself.” Many China watchers sent the story to me, suggesting that it proved that China was much closer to the U.S. than it appeared on the issue of not wanting to build AGI/RSI. But I wanted to know more about how Chinese labs and the state really thought about AI safety, so I asked my sources about it.
Jordan Schneider, responding in part to the above Times story:
Obviously, an authoritarian government doesn’t want autonomous AI agents running around and taking control away from the state. But China is Leninist, and there is a strain of Leninist thought that isn’t afraid of running the economy as a machine. I think Khrushchev and Brezhnev would have loved to entrust Soviet economic planning to a superintelligent machine.
There is the military argument: We all know AI can be a powerful weapon. There is the surveillance case: For the CCP, there is something inherently appealing about having an AI-powered panopticon. There is the economic argument. The Chinese economy has been stuck at slow growth for years. The bet Xi has made in the future of the economy is on emerging technologies. If advanced AI holds the potential for China to have widespread productivity growth, he wants that. So, as a literal weapon, as a potential driver of productivity, as a tool for surveillance, and as a point of international prestige, AI is, and will continue to be, very appealing to Xi.
Paul Triolo, partner at DGA-Albright Stonebridge Group:
The Chinese safety community cares about AI safety, and is increasingly sophisticated in its assessment of risks from advanced AI. Chinese AI companies increasingly care about safety, as they want to make sure that their models are not misused by bad actors. The Chinese government has been slow to mandate AI safety teams and testing at leading AI companies, in part because of the desire not to overregulate the companies and ensure they are competitive with US leaders. At the same time, the Chinese government would have a hard time mandating safety testing when Chinese companies are already short of advanced compute capacity. So at a minimum, the Chinese government should seek to mandate AI companies have AI safety testing teams, and then possibly provide a national testing platform to ensure there is sufficient compute for safety testing.
Matt Sheehan:
China’s AI safety community is small. They have some influential connections and allies. But it’s not like the U.S., for several reasons. First, in China, there is a sense that the party is going to make the big decisions on national security, so the companies feel less personal responsibility [to fixate on safety]. Second, the Chinese labs are behind the frontier labs, so they’re not yet seeing the same capabilities that are scaring Anthropic and OpenAI researchers. Finally, while the Chinese government imposes a lot of requirements on AI companies, the overwhelming majority of those requirements regulate political speech rather than other kinds of risk. It’s mostly about controlling information, not regulating frontier-AI safety.
Kyle Chan, fellow at the Brookings Institution:
A lot of this comes down to one question: What is your timeline for takeoff? Most Chinese researchers don’t think that ASI/RSI is right around the corner. Many American AI researchers do. It’s hard to argue that AI safety around RSI should be a priority if you don’t think recursive self-improvement is imminent. China is not likely to want to slow down, but it’s largely because they don’t see AI as an imminent existential threat.
Xiao Junyong, law professor at the Beijing Institute of Technology, writing in the People’s Daily (a CCP official newspaper):
On close examination, the operating logic behind this “Build American AI” push is directly inherited from the “China threat” theory playbook of the past.
First, set up a target. Standardized talking-point templates are distributed, casting Chinese AI as the “principal external threat,” claiming it will harvest the public’s information and squeeze Americans out of their jobs.
Second, deflect internal contradictions. America’s leading AI firms are facing real pressures — high energy-consumption costs, controversies over job displacement — and by hyping up an external threat they steer the focus of domestic public opinion toward China.
Third, secure resources. These companies are priming the atmosphere ahead of upcoming elections and pushing Congress to increase federal subsidies and loosen regulation of the sector. This is cut from exactly the same cloth as the old drumbeating about the “China military threat” to drive up defense spending, or politicians hyping the “China Shock” to push through trade protectionism.
Matt Sheehan:
China thinks the frontier labs are insincere, calling for regulating AI while also speeding up. They see Dario saying we should do everything to contain China but also that we must pace the frontier. China only heard the first part of that sentence, and it’s muddied the waters around the case for pacing. On the US side, we have our own version of conspiratorial thinking. One example is in the Dario letter, about how if China gets to ASI first they’ll impose authoritarianism uniformly across the globe. That is a huge extrapolation.