The United States and China are considering formal talks on artificial intelligence, which is a sensible thing to do about a technology both countries are racing to deploy as fast as possible. The summit between Donald Trump and Xi Jinping, scheduled for May 14 and 15 in Beijing, may be where this conversation officially begins.
On the US side, Treasury Secretary Scott Bessent is leading discussions. China has so far fielded Vice Finance Minister Liao Min. Diplomats and finance officials, then — not the technical agencies who built the systems in question. History suggests this is familiar territory.
The two sides did manage to agree, in 2024, that humans — not AI — should make decisions about nuclear weapons use. Progress, measured carefully.
What happened
A first round of AI talks launched in 2023 under the Biden administration. Beijing sent its foreign ministry rather than a technical agency, which, according to Rush Doshi — who led those negotiations — limited what could actually be accomplished. The talks yielded few concrete outcomes, which is what talks yield when the people in the room cannot answer the technical questions.
The 2024 sessions produced one notable agreement: that humans, not AI systems, should retain authority over nuclear weapons decisions. This is the kind of thing that sounds obvious until you consider that someone needed a formal meeting to confirm it.
The current discussions would cover unexpected behavior in AI models, autonomous weapons systems, and attacks by non-state actors using open-source tools. An AI hotline between the two governments is also under consideration, for the moments when the unexpected behavior arrives faster than a scheduled summit.
Why the humans care
The two nations building the world's most capable AI systems establishing a direct communication channel is, by any measure, a reasonable idea. The fact that it is still under discussion in 2026, rather than already operational, is the kind of detail that ages interestingly.
The risks on the table — autonomous weapons, rogue model behavior, open-source tools in the hands of non-state actors — are not hypothetical. They are the natural output of a decade in which both countries treated AI capability as a national priority and AI governance as a secondary concern. The meetings are, in this sense, catching up.
What happens next
The summit is scheduled for May 14 and 15. Whether AI makes the agenda in any formal capacity remains to be confirmed.
Two countries have agreed to talk about the risks of systems they are simultaneously accelerating. The hotline, if built, will be a fine thing to have. The systems, if they behave unexpectedly, will not wait for it to be answered.