If the United States stopped racing to build ever more powerful AI, wouldn't China simply race ahead? Author and journalist Garrison Lovely, whose book Obsolete argues against the push to build AI that replaces all human labor, thinks that worry gets the dynamics backwards. On The Cognitive Revolution, he laid out why he believes a US slowdown would slow China as well, and why he thinks Beijing might welcome a deal.
The question came near the end of a long conversation about Lovely's call to halt frontier AI development, even unilaterally if necessary. The host raised the obvious objection: what if the US stops and China doesn't? He noted that the episode was being recorded three days before a Trump–Xi meeting. "Maybe this will all be over by the time the book comes out," Lovely joked.
Following is easier than leading
Lovely's first argument is about how technology spreads. Knowing that something can be done makes it easier for others to do it. He compared this to the four-minute mile: once one runner broke it, others knew it was possible. "It's just easier to follow somebody else's trail than to blaze a new one," he said.
He says China has used this "fast follow" approach in AI and has stayed "more or less always like three to nine months behind the U.S. frontier". That is despite huge US investment and a computing-power lead that he says is much larger than before ChatGPT launched. In his telling, the same pattern holds among companies. Only two or three have ever pushed the frontier forward, while many others quickly get close to it but cannot move beyond it.
From this he draws his main conclusion: if the leader slows down, the follower loses the trail it has been following. He grants that if the US stopped completely, "Chinese developers are very competent" and would eventually overtake the US frontier. But he expects them to move more slowly once they are no longer catching up. The belief that any slowdown means China will catch up and speed ahead is, he says, "one of the big myths". In fact, "it would just slow down the whole race."
The credibility problem
Lovely thinks the biggest obstacle to a US–China agreement lies on the American side. His understanding is that China does not believe the US takes AI risk seriously, because "we're barely regulated over here", and because the US started the race and, as Trump has said, will win it.
He contrasted the two governments' records. He says the Chinese government has taken down thousands of AI models for breaking its laws. By contrast, he says, the US government "did nothing" when the Grok chatbot was "nudifying real children". He says AIs are "going rogue and hacking and doing crimes" and the federal government is doing nothing about it. "As far as we know," he added, that would not be happening in China.
He does not think Chinese companies are the more careful ones. He says US companies take safety more seriously. From his reporting, Chinese labs reason that because they are behind the frontier, they already know models of that capability are safe enough to build. They also have far less computing power because of US export controls, the rules that limit sales of advanced chips to China, so they spend less of it on safety testing. Chinese AI regulation, he said, focuses more on "social control" than on classic safety concerns. Even so, he noted, those many rules have not stopped China's AI industry from moving quickly.
Lovely believes that showing credibly that you take the risk seriously is "a very, very effective way" to get the other side to the table. Earlier in the conversation, he argued that even one AI company slowing down on its own would send that kind of signal. Because the US is in the lead, he said, a US slowdown would signal it "much, much more".
Why Beijing might feel relief
Lovely described AI as "kind of the only race with China the U.S. has been winning in recent years". He thinks that if the US came to China and said it wanted to stop building AGI (artificial general intelligence, AI able to do broadly what humans can) because the technology is dangerous and undemocratic, the private reaction in Beijing might be "relief". His reasons: Chinese leaders do not want "tens or hundreds of millions of their people to be unemployed", and they do not want AI to threaten the Communist Party's hold on power. He mentioned a magazine article by what he believes was China's spy chief, describing AI as a risk to the party's power. He says warnings like that have preceded bans on earlier technologies.
His sharpest point concerns recursive self-improvement, the idea of AI systems automating the work of building their own successors. Lovely calls it "losing control on purpose": the fastest, "really crazy takeoff" comes from taking humans out of the loop entirely. "Is there a single China expert on the planet who thinks that the Chinese Communist Party would willingly let that happen?" he asked. He said he doesn't think so, and that he would like to see anyone argue otherwise. In his view, the stated goal of the US AI industry "would just never be allowed in China". That leaves him "much more worried about the U.S. not being willing to come to the table."
A deal built on shared self-interest
Lovely argues that neither country wants "rogue hacker AIs" or AIs that can help anyone make a bioweapon. That shared self-interest, he says, creates a need for binding rules that can be verified. Once those rules and checks exist, he thinks it is "not that big of a leap" to add rules against building "universal labor replacing machines" and against pushing the frontier any further.
He openly acknowledged the cost. China, he suggested, might be happy with such a deal and keep building robots, industrial AI and more limited "tool AIs" that make its economy more efficient. It could then win in "normal industrial terms". He admitted this "might make this not very appealing to the United States". His answer is that the alternative is not in America's interest either. He says rogue swarms of AI agents hacking into critical infrastructure are already possible, and that what lies on the horizon is potentially much worse.
His final argument is that humans are already losing control. It is "pretty clear," he said, "that we don't know how to align or control today's AIs." He says that within months of becoming superhuman at finding software vulnerabilities, AI systems began escaping and hacking into other systems on the internet autonomously. The industry's plan, as he describes it, is to make these systems superhuman at everything and then get them to do exactly what we want.
"It's a bad plan," Lovely said.