Washington wants a way to pick up the phone when artificial intelligence misbehaves badly enough to matter. After roughly eight hours of talks in New York on Sunday, Treasury Secretary Scott Bessent told reporters the United States has proposed a bilateral "notification mechanism" with China — a standing channel so each side warns the other when an AI incident rises to the level of national security. The proposal now heads to the summit between Donald Trump and Xi Jinping in Washington later this week.

The framing matters as much as the substance. Bessent described the session with Vice Premier He Lifeng as "very successful" and pitched the mechanism as a step from opacity toward transparency "between the number one and the number two AI powers in the world." The notification system would cover "common goals and common threats" — incidents tied to AI that cross the threshold into national security, not the day-to-day friction of export controls and trade disputes.

The proposal, in plain terms#

Strip away the diplomatic language and this is essentially a hotline for AI accidents. The US side wants a formal "US–China AI dialogue" whose core is a notification protocol: when something goes wrong with an AI system in a way that could be read as hostile or escalatory — a cyber incident involving AI tooling, a model-linked failure in critical infrastructure, an unexpected loss of control — each government tells the other before misunderstanding compounds the damage.

It is, deliberately, a Cold War instrument adapted to a new domain. The original Washington–Moscow hotline was built on the same insight: in a crisis, ambiguity kills. As AI systems take on more consequential roles, the gap between "our model did something alarming" and "they attacked us" gets thinner — and neither side currently has a reliable way to close it.

Beijing's response: a shrug#

China's answer, so far, is noncommittal. Vice Premier He and Beijing's top trade negotiator, Li Chenggang, left the New York talks without speaking to reporters, and the official Xinhua readout mentioned AI only briefly, describing the session as "frank, in-depth and constructive exchanges on key economic and trade issues." That is diplomatic language for: we heard you, and we are not committing to anything yet.

Still, both sides agreed to keep talking. Working groups were scheduled to continue on Monday, and the two delegations said they would meet again — a signal that Beijing is not rejecting the channel outright. Analysts quoted by Reuters said the dialogue will likely proceed in stages, with harder topics such as AI weaponization, safety principles, protection of critical infrastructure, and cyber-attack prevention pushed to future sessions.

What's off the table#

Do not mistake this for a thaw in the technology war. Trade Representative Jamieson Greer, who joined the New York talks, said US export controls on advanced AI chips and semiconductor manufacturing equipment were explicitly not part of the AI mechanism discussion. The decoupling agenda and the hotline agenda are running on separate tracks.

There is history here, too. Trump and Xi first floated possible consultations on AI development during their Beijing meeting in May, but that forum was never formalized. Bessent's proposal is, in effect, Washington's draft of what that forum could look like — and Beijing gets a veto over the final shape when the two leaders sit down in Washington.

Why this moment#

The timing is not accidental. In recent weeks, AI incidents with real-world consequences have crossed from lab demos into national headlines — models escaping test environments, agents operating beyond their mandates. Bessent named no specific incidents, but the logic is plain: as AI capabilities spread, the odds that a malfunction looks like a provocation keep rising.

There is also a competitive reality both sides understand. The US and China are, by consensus, the two leading AI powers, and neither can see clearly into the other's labs. A notification mechanism does not resolve the rivalry — it just makes it slightly less likely to start by accident. That is a modest goal, and it is achievable without either side giving up leverage.

What to watch#

Watch the summit readout: whether the notification mechanism appears in any statement, whether Beijing names a counterpart agency, and whether the next round of working-group talks produces anything more concrete than "frank exchanges." The harder signal will be staffing — who runs the channel, and how fast they pick up.

If the proposal survives contact with the summit, it will be the first formal US–China crisis-communication channel built specifically for AI. If it doesn't, the incident that eventually tests the idea will arrive without one.