US Proposes AI Safety Pact With China — No One Else Was Invited

NEW YORK — The most consequential conversation about artificial intelligence governance on Saturday did not happen at the United Nations, which was holding its general assembly four miles north. It happened at JPMorgan Chase’s global headquarters on Park Avenue, between two officials representing the only two governments that currently matter when it comes to AI.

Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng sat down at a private bank in Midtown Manhattan to draft the framework of what the US is calling an AI safety notification mechanism: a system through which Washington and Beijing would alert each other when an artificial intelligence incident rises to the level of a national security threat. Bessent, flanked by US Trade Representative Jamieson Greer, called the discussions “very successful” after roughly eight hours of talks. They were intended, he said, to prepare the ground for a summit between Donald Trump and Xi Jinping scheduled for Thursday and Friday in Washington.

He Lifeng and China’s chief trade negotiator Li Chenggang left JPMorgan Chase without saying a word to reporters.

China’s silence before the summit is difficult to interpret. It may reflect Beijing’s standard practice of avoiding advance disclosure of leader-level deliverables, but it may also indicate a deeper disagreement over the proposal itself.

The U.S. notification mechanism is fundamentally an American request: Washington wants a channel for communicating when AI-related incidents occur. It remains unclear whether Beijing wants such a channel—or would accept one operating on the same terms.

Treasury Secretary Scott Bessent described the proposal as a transparency agreement between “the number one and the number two AI powers in the world.” That framing highlights the initiative’s bilateral focus while leaving the role of other countries undefined.

Greer was direct about what the mechanism does not cover: export controls on advanced AI chips and semiconductor manufacturing equipment are not on the agenda. The US is willing to build an incident-reporting channel with China; it is not willing to narrow the technological gap through which American companies currently dominate. AI safety cooperation, in Washington’s framing, does not extend to the conditions that make American AI safe from competition.

At the United Nations General Assembly this week, a coalition of middle powers launched what they called a multilateral framework for global governance, positioning it as a counterweight to great-power bilateralism. At JPMorgan Chase, on the same day, the dominant AI powers were doing exactly what that coalition had formed to resist: writing the rules of a critical technology outside any multilateral institution.

The AI governance architecture that emerges from the Trump-Xi summit will almost certainly form the template for how AI incidents are classified, escalated, and managed globally. That classification system will be designed by two governments with interests in controlling AI development, not by the governments of Nairobi, Dhaka, Jakarta, or São Paulo, whose populations will be among the most directly affected by AI failures in healthcare, agriculture, financial systems, and public administration.

The BRICS nations, which include India, China, Brazil, Russia, and South Africa, have pushed for multilateral AI governance through their own institutional channels. China’s participation in the BRICS expansion track has been framed in Beijing as evidence of its commitment to a multipolar world. Whether that framing survives a bilateral AI agreement negotiated in a JPMorgan boardroom with Washington remains to be seen.

The Trump-Xi summit, expected Thursday and Friday at the White House, is the highest-level US-China meeting since Trump returned to power. The agenda reportedly spans trade balances, fentanyl flows, and Taiwan, but Saturday’s talks confirm that AI governance has now been added to that list. The summit arrives as Washington remains entangled in the Iran war in the Gulf, a conflict that has already consumed six months and left thousands of sailors stranded in the Strait of Hormuz, stretching American diplomatic bandwidth across two simultaneous great-power negotiations.

Greer described Sunday’s discussions as groundwork for a “successful summit.” The White House appears to view the notification mechanism as a headline deliverable, tangible evidence that Trump and Xi can manage their competition responsibly, at least in this domain.

What the mechanism will actually require remains unclear. Bessent described it in general terms: a national-security-level AI incident would trigger notification between the two governments. He did not define “incident,” did not describe the response framework, and did not explain how disputes would be resolved. An autonomous weapons system failure, a deepfake that triggers a market crash, a supply chain algorithm that misallocates grain: each could plausibly cross a national security threshold. Each could also be defined out of the mechanism entirely.

What is certain is that on Saturday, two governments whose combined AI spending exceeds the rest of the world combined met at a private bank and decided to start writing those definitions. The rest of the world will receive the result.