NEW YORK — The world’s two dominant artificial intelligence powers agreed Sunday to establish a formal channel for warning each other when their AI systems generate incidents serious enough to affect national security — a first-of-its-kind arrangement reached after eight hours of talks at JPMorgan Chase’s Manhattan headquarters between Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng.
The agreement, announced by Bessent after the talks concluded late Sunday, is the most substantive AI governance step the United States and China have taken since competition between their AI industries began escalating in 2022. Both sides want “a shared vision of common goals and common threats,” Bessent told reporters. The proposed mechanism would create a notification system allowing each government to alert the other when an AI-related incident rises to the level of national security concern — a category broad enough to encompass everything from autonomous weapons behavior to infrastructure penetration, but specific enough to exclude routine commercial AI failures.
What the dialogue does not cover is as informative as what it does. Export controls on advanced semiconductors and AI chip manufacturing equipment remain entirely outside the arrangement’s scope. The two countries are each continuing to develop the most capable AI systems in the world, and Bessent’s framing — “common threats” — was careful not to describe Chinese AI development as itself a threat, a distinction Washington has not always maintained. The dialogue addresses behavior and incidents, not the pace or direction of development.
The eight-hour session at JPMorgan’s midtown offices also included US Trade Representative Jamieson Greer on the American side and Li Chenggang, China’s top trade negotiator, as part of He Lifeng’s delegation. Three agenda items drove the meeting: the status of the Busan trade truce, which expires November 10; China’s underdelivery on rare-earth export commitments it made under the truce; and the AI guardrails framework. None of the three reached formal conclusion by the time the talks wrapped Sunday evening. A working group is scheduled to continue Monday before the Trump-Xi summit at the White House on Thursday.
Li Chenggang described the discussions, in a separate statement to reporters, as being conducted in “a good atmosphere” — diplomatic language that confirmed the talks did not break down without committing to the existence of substantive agreement. Al Jazeera reported that neither side had addressed how quickly notifications under the proposed AI mechanism would need to be issued, or what the receiving country would be obligated to do upon receiving one.
On trade, the outlines of what Thursday’s summit might produce have become clearer. American negotiators are focused on Chinese purchases of US energy, agricultural goods, and potentially medical devices. Beijing’s delegation is protecting high-value exports while offering consumer goods and low-technology items as China’s side of any tariff reduction. Discussions have settled on a mutual framework covering $30 billion of products on each side — a figure both parties have confirmed publicly. Whether those discussions crystallise into a formal announcement Thursday is the question markets have been pricing since Xi landed in Washington on Sunday; the November 10 expiration of the Busan truce is the mechanism applying pressure on both sides to conclude something.
Xi arrived in Washington on Sunday under conditions that framed the summit as partly driven by the tariff truce’s expiration — a clock neither side set deliberately but both are now running against.
The rare-earth question is the trade dispute most likely to provoke friction Thursday. China had committed, under the Busan framework, to restore export volumes of rare-earth elements and processed minerals critical to semiconductor manufacturing and defence applications. American officials say Beijing has delivered a fraction of the committed volumes. Beijing has not publicly contested the assessment but has also not offered a timeline for delivery. A resolution — even a partial one — on rare earths would represent more concrete progress than the AI dialogue, which remains a framework document without implementation rules.
The difficulty of implementation has shadowed every US-China agreement of this kind. The countries established a military communications hotline in 2014 that was allowed to fall into disuse. They reached understandings on economic espionage in 2015 that lasted under a year. What makes the AI dialogue different, if anything does, is the specificity of the problem it addresses: both governments have acknowledged, at least in background conversations, that their AI systems have already generated incidents neither side fully controlled. Whether acknowledging that risk in a bilateral framework actually reduces it, or simply gives both governments a way to describe the problem without addressing its source, will determine whether Sunday’s agreement matters beyond this week.
