Bessent says US‑China talks on advanced AI models will move forward at Shenzhen meeting

US Treasury Secretary Scott Bessent said on Tuesday that next week's Trump-Xi meeting in Shenzhen will advance a new US-China dialogue on advanced AI. He also said the two sides have agreed to set up a direct channel for alerting each other to serious incidents, Reuters reported.

The remarks matter because they put dates and mechanisms behind what had been loose talk of AI cooperation between the two main developers of frontier models.

What Bessent said

Bessent spoke in New York in prepared remarks to the National Committee on US-China Relations. Reuters said he will travel to China next week for President Donald Trump's third meeting this year with President Xi Jinping.

That meeting will take place on the sidelines of the Nov. 18-19 Asia-Pacific Economic Cooperation summit. Bessent said it will advance a bilateral dialogue on "superintelligence", the term the Trump administration prefers.

He also said the sides expect to exchange best practices for evaluating and testing advanced models before release. The dialogue is meant to guard against shared risks and "prevent a misunderstanding between our two systems", Reuters quoted him as saying.

He wasn't selling a partnership. In the Treasury's published text of the speech, he said the two countries "should harbor no illusions" about their differences. On trade, he said the US objective is to "de-risk, not to decouple".

The trade side

Reuters said Bessent also pointed to a "Board of Trade" that would set mutual tariff cuts on $30 billion of non-sensitive goods in each direction. He cited China's pledged purchases of US farm goods too. A similar "Board of Investment" would handle market access for American companies in China.

How the AI track got here

The AI dialogue has been built in stages over six months. In May, Bessent told CNBC that the two countries would set up a protocol to keep non-state actors from getting hold of powerful models. He added that Washington can hold such talks because "we are in the lead".

On Sept. 20, Bessent and Chinese Vice Premier He Lifeng met in New York. CNN reported that the US side proposed a notification mechanism covering AI incidents that reach a national-security level. US Trade Representative Jamieson Greer said export controls on AI chips and chipmaking equipment weren't on the agenda for those talks.

Before that meeting, Bessent said he expected the discussions to cover "both open- and closed-weight models", according to Reuters via CNBC. Open-weight models are systems whose core components are public, so users can download and fine-tune them.

Trump and Xi then met in Washington on Sept. 24, The Diplomat noted. The Asia Society's Policy Institute wrote that the two leaders agreed to establish an AI dialogue, with the next round expected in Shenzhen.

Where the two sides differ

The sides don't even use the same name. The Asia Society pointed out that Beijing's readout calls it an "AI Dialogue", while Washington's calls it a "Super Intelligence Dialogue".

Reuters added that Trump agreed to the alert system in September but also said he was reluctant to add new guardrails because he wanted to preserve US leadership.

There's friction on the models themselves, too. In July, Bessent told Fox Business the administration would look into whether Chinese models were distilled from American ones. He said the US has the ability to sanction overseas models found to be stealing. CNBC also reported that Axios has said Washington is considering limits on US use of Chinese open-weight models.

The stakes for developers and users

For AI companies, the immediate effect is limited. As far as I can tell, no new rules on model releases or chip exports came out of Tuesday's remarks. Reuters reported in July that the administration was discussing voluntary safety standards with US AI firms but hadn't announced regulatory measures.

What changes is the channel. If an incident reaches national-security level, the plan is for each government to have a direct line to the other. Bessent's description doesn't say what would count as an incident or who would run the line.

My view: an incident hotline is a modest step, and probably the easiest piece for both sides to agree on. Pre-release testing is harder, because it touches the same competition Bessent says the US intends to win.

What happens next

Trump and Xi are due to meet in Shenzhen on the sidelines of the Nov. 18-19 APEC summit. The Asia Society reported that both leaders also plan to attend the G20 summit in Miami in December, which would make four meetings this year.

Honestly, I'd wait for the Shenzhen readouts before reading much into the dialogue. If Washington and Beijing publish different descriptions of what they agreed, as they did in September, that will tell you how much common ground there really is.