How Trump and Xi Can Do AI Safety
ChinaTalk Jay Kimmel
Trump and Xi are set to talk AI safety, but the agenda is still fuzzy. The big risk: another empty U.S.-China dialogue; the best chance is a narrow deal on shared dangers.
Based on reporting by ChinaTalk, Jay Kimmel — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
As the next Trump-Xi summit gets closer, the AI safety channel the two sides announced in May still looks thin on details. That’s a familiar problem. U.S.-China talks have a habit of getting swallowed by bigger fights, or of turning into long grievances sessions that produce very little anyone can use.
The author of the piece, Jay Kimmel, argues that the first mistake would be aiming too high. He points to past breakdowns after Pelosi’s 2022 Taiwan visit, the EP-3 collision, the NATO bombing of China’s embassy in Belgrade, and even the 2023 spy balloon episode, when a crisis hotline existed on paper but the PLA did not take the call. The lesson is blunt: a mechanism only matters if both sides are willing to pick up.
So the case here is for starting small and technical. The strongest candidates, based on an expert survey, were nuclear risk, CBRNe threats, biosecurity controls on models, risks from non-state actors, and use of new models to close vulnerabilities in open source software. The U.S.-China experts were most cautious about feasibility, with only nuclear risk and open-source software vulnerabilities clearing the 50% mark for a productive dialogue. The Track II group was more optimistic, and saw the issues as more urgent.
Kimmel’s wider point is that governments shouldn’t try to do this alone. Both Beijing and Washington are still assembling their AI policy circles, while companies, technical bodies, scientists, think tanks, and AI safety groups already have useful pieces of the puzzle. Public comment on what should go into talks, what warning signs matter, and which industry practices might be worth joint endorsement would be an easy early win.
He also doesn’t pretend the politics are clean. Some U.S. officials see China itself as the main risk, and Beijing’s language around being “secure and controllable” won’t mean the same thing to everyone. But the author still thinks there’s a narrow lane for cooperation where both sides share an interest in avoiding catastrophe. That lane begins with information sharing, then, if the mood holds, something sturdier.
My take — AI-written commentary, not fact-checked reporting
The fantasy here is that AI safety talks are a magical escape hatch from the rest of U.S.-China relations. They’re not. The only version that survives contact with reality is boring, narrow, and technical — which is exactly why it has a chance. Anything grander will probably end up as diplomatic wallpaper, with better adjectives.
Read more about this at: ChinaTalk
