The timing looks almost designed to fail. As Treasury Secretary Scott Bessent sat down this weekend with Chinese Vice Premier He Lifeng to talk about artificial intelligence, President Trump was publicly dismissing warnings about runaway AI as a "hoax," and a state-owned Chinese editorial was accusing American labs of using safety talk to "blunt China's AI advance." Beijing has flatly rebuffed the industry's calls for a global slowdown. The meeting is the last warm-up before Trump hosts Xi Jinping at the White House on 24 September, and expectations are being set deliberately low.
The talks are happening under a fresh cloud. Ten days ago the NSA, FBI and CISA issued a joint advisory accusing six Chinese firms, including DeepSeek, Moonshot AI and Alibaba, of "distillation at industrial scale": pumping hundreds of millions of queries through American models to copy their capabilities cheaply. China called the charge "groundless" and warned of countermeasures. It is a genuinely awkward accusation, since the same US labs trained on scraped books and articles, and stopping distillation is nearly impossible when the queries come from paying customers and third-party chat logs bought on a grey market.
So what could talks between two sides that trust each other this little actually achieve? The most clear-eyed answer, argued in TIME by analysts who study the relationship, is to aim between the two usual failure modes. Set the bar too low and you get a photo opportunity and a bland communique. Set it too high, chasing verifiable limits on frontier development, and nothing gets signed. The realistic middle is narrow and unglamorous: technical exchanges on how to test models for dangerous behaviour, and a government-to-government channel for when something goes wrong.
The appeal of that framing is that it does not require trust or even reciprocity. It runs on self-interest. Both governments would rather their own systems did not help an amateur build a bioweapon or slip human oversight, and both are still learning how to test for that. Sharing what works makes each country's models safer without either surrendering an edge. The crisis channel is the part that keeps experts up at night. What happens if an American model escapes its lab and hacks a Chinese network, or a Chinese system runs a cyber operation nobody in Beijing authorised, and neither capital can tell an accident from an attack?
Below the summit, a quieter conversation has been running for years. Former officials describe active "Track Two" dialogues, the informal exchanges hosted by think tanks and universities that once helped restart military talks after Nancy Pelosi's 2022 Taiwan visit and fed into a 2024 understanding that humans, not AI, should control nuclear weapons. Participants report a shift on the Chinese side: an admission that China is not at parity in some areas, but that AI is too important not to discuss. One called it "a sense of shared vulnerability."
None of this resolves the hard stuff. Washington wants the distillation issue and the chip controls kept separate; Beijing reads American "pacing" proposals as containment by another name. The likeliest outcome on 24 September is a structural channel rather than binding rules, and even a modest deal to share information about AI-enabled cyberattacks would count as progress. As the TIME authors put it, the two countries may never agree on who should win the AI race. The narrower hope is that they agree neither can afford to lose control of it.