Sign up for the Slatest to get the most insightful analysis, criticism, and advice out there, delivered to your inbox daily.
“If they’re going to be killer robots, I’d rather they be American killer robots and not Chinese killer robots,” Ted Cruz told Politico in early September. It was the most concise version of the argument that the Republicans running the country have made often against artificial intelligence regulation. They equate regulation with an American slowdown that would put Chinese A.I. developers ahead. “WHOEVER WINS AI, WINS!” President Donald Trump posted on Sept. 14. The only guardrail we needed was “a STRONG AND SMART (High IQ!) PRESIDENT.” It sounds like we’re all good, then, no regulation or negotiation required.
You may be of the mind that to hand-wring right now about A.I.’s hypothetical potential is to give in to disaster-based marketing that benefits the major A.I. labs and their investors. But the most powerful people in the country are reaching the same conclusion—that we shouldn’t get too precious about regulating A.I.—for the opposite reason. They think A.I. could be so important that overpolicing American A.I. companies would hand a devastating geopolitical edge to China.
For the past few weeks, I’ve been emailing and calling people who spend a lot of time on either A.I. research or the U.S.–China relationship. They make a convincing case that the U.S. government can’t prevent A.I. development from going to unsavory places. But the case that China will take it to those places, and so the United States needs to keep pushing on A.I. to make sure it doesn’t lose to China, is less persuasive.
“There are real asymmetries one can worry about between countries in so-called ‘A.I.’ technologies,” Graham Webster, a Stanford University research scholar who studies Chinese technology, told me in an email. “But the idea of a single A.I. race is, I believe, a distraction from what’s actually going on.”
Even if you take the hawkish view, there are ways in which it makes sense that the U.S. could slow down Chinese firms. A long-standing complaint of the American A.I. industry is that Chinese model makers are “distilling” closed-source models from Anthropic and OpenAI, spinning up thousands of fake accounts to harvest the models’ answers and train their own systems on them. Anthropic’s CEO, Dario Amodei, has been public about his frustration with Chinese developers trying to rip off his tech.
“I’ve heard people make this argument in an almost snarky way,” Adam Thierer, a senior fellow at the R Street Institute, a center-right think tank in D.C., told me. “ ‘Look, if Dario and Co. really want to slow down the looming threat of existential A.I. risks, then they should just slow down their own models and, by extension, that will stop the distillation by Chinese labs of our models, and the whole world gets safer.’ You know, this is a big gamble to make.”
Thierer made the point that China has made big strides in computational sciences and that its A.I. development can’t possibly be explained solely by intellectual property theft. “The Chinese have some of the best and brightest in the world, and their universities are cranking out incredible work and papers,” he said. “They out-publish us on this front.” In other words, even if China is copying Claude, limiting Claude may not slow down China.
But it’s also not likely that competitive secrets are flowing in only one direction. Chinese models tend to be open-weight, meaning that the public (and, certainly, the biggest American companies) can look under their hoods. “It’s something that’s often overlooked,” Thierer told me. “I think our labs probably are taking a hard look at Chinese models.”
If China is racing, it seems like a different race from the one Cruz describes: not to build the single most powerful system and keep it, but to get its models into as many hands as possible. China doesn’t seem worried about many of those hands not even being Chinese. An open-weight model offers the same degree of access to everyone, not just Chinese developers and scientists. Xi Jinping is trying to create a whole sphere of open-source A.I. in conjunction with the rest of the BRICS countries. China’s push for more open-source models is in opposition to the United States, in that the country doesn’t want all of the world’s A.I. products to be built with a few American models. But China isn’t hoarding its models for itself in a one-vs.-one race either. Its open-source models on Hugging Face, an A.I. platform, got more downloads over the past year than its American peers did, according to a report the platform published in March.
Trump’s view of Chinese regulation is that it’s a free-for-all. The president posted this month that he had interpreted Xi’s announcement of that initiative to mean that China “will be doing absolutely nothing to stand in the way of A.I., or its future.” Yet China already does more national A.I. regulation than we do. It just does it by policing how A.I. reaches the public, rather than how it’s developed. “To say that China’s not regulating, I think, is incorrect,” Sarah Kreps, the director of Cornell University’s Tech Policy Institute, told me. “It’s just that they’re not interested in holding back these open-weight models.”
Kwan Yee Ng, head of international governance at Beijing-based Concordia AI, which calls itself a safety consultancy, told me in an email that Chinese developers already need to pass security checks for their algorithms, that users need to register with real names, and that the country has “strict content labeling standards for A.I.-generated material.” The country is also now trying to crack down on the use of A.I. bots as “companions.” She said China was never going full speed ahead on A.I.
“The way many in China see it, China has been cautious from the start,” she wrote.
The Chinese government has put its thumb on the scale of the development of A.I. models at a high level. In 2023 the state internet regulator said that “the provision and use of generative artificial intelligence services should adhere to the core socialist values.” Beyond that, China’s restrictions have not been on the development of the models but on their use cases, like if they can be used for “companionship” and how their outputs are labeled.
That sounds like a country that is bullish on A.I.’s potential but has already taken more steps than the United States has to address the tech’s very present, nontheoretical downsides. A country thinking about A.I. in that way seems like one you could negotiate with if you were really concerned about “killer robots.” It’s questionable how useful it would be for Trump to put Jared Kushner and Steve Witkoff on this case unless the goal was for the Trump family to get an equity stake in DeepSeek in exchange for American policy concessions. Actual diplomats might have better luck. In fairness, they also might not.
“It took 17 years and a Cuban missile crisis to force the U.S. and Soviet Union to the table to conclude the Partial Test Ban Treaty in 1963,” Kreps said. “You [had] already demonstrated the super kinetic and destructive capability of nuclear weapons, and it still took 17 years and this major crisis.”
Michael Horowitz, a University of Pennsylvania political scientist who served as deputy assistant secretary of defense for emerging capabilities during the Biden administration, told me in an email that A.I. is a general-purpose technology, so “the economic incentives for development are incredibly strong.” He said regulations on this kind of tech “tend to be standard setting rather than active constraints.” That gives China a straightforward economic incentive to keep going, even if the U.S. did want to deal.
“A.I. is not a nuclear weapon,” Horowitz said.
A sweeping U.S.-China A.I. accord, in this political and financial economy, may indeed be the longest of long shots. But China has other incentives not to want A.I. to get out of hand, and they fly in the face of the idea that any U.S. slowdown would automatically leave China to sprint into the future alone. A sci-fi-style loss of control of A.I. would stink for lots of governments, but it might stink a bit more for an authoritarian government that censors things all the time and is uncomfortable without a firm grip on the public.
“I think it makes a great deal of sense in theory that China would want to crack down on A.I. technologies and algorithmic speech,” Thierer said, “because they have a history of cracking down on digital speech and the internet more generally.” The head of China’s spy services sounds worried about A.I.’s potential to disrupt Communist Party rule. Welcome to the pretzel of A.I. politics, where a person could make a moral argument against model development that aligns with the Chinese government’s desire to crush dissent.
That doesn’t sound all that much like a “race,” though. Webster, the Stanford researcher, unpacked why that framing is so unhelpful in a recent Substack article. Say that Google achieves some A.I. breakthrough that vaults Waymo to market dominance in the autonomous ride-share space. “The US government doesn’t own Waymo, nor do the American people,” Webster wrote. “Some shareholders, who are spread around the world, might gain. Some riders enjoy the service. If it is wildly successful, already-struggling gig workers likely will not enjoy the increased squeeze on their finances. And Chinese life proceeds largely unaffected.” In that sense, an American “win” would not do much for you, an American. The same could be true for any given Chinese consumer in the areas where Chinese firms “win” with specific products.
What of a “loss,” though? I asked Kreps what the most well-founded fear was in the event that Chinese models became dominant. Would China gate them off from Americans, seizing a productivity advantage for itself? Would Americans use them en masse, creating a data privacy problem? Would the Chinese models, tuned to accommodate the party line, turn us all into frothing communists?
“I think it’s all of the above, honestly,” she said.
