White House denies a mid-September AI meeting as sources describe Bessent-led talks

The United States and China are preparing a mid-September dialogue on AI safety, two people briefed on the planning told Reuters, in what would be the first official talks devoted only to artificial intelligence since Donald Trump returned to the White House. U.S. Treasury Secretary Scott Bessent would lead the American side. A White House official said there is currently no planned AI-related meeting in mid-September. Trump and Xi Jinping are due to meet in Washington on 24 September.

Those two statements can sit in the same week. A denied meeting on the calendar is not the same thing as denied talks. Reuters published the mid-September window on 4 September. The White House line followed. Neither side has issued a joint notice. On the Chinese side, one of the people said Vice Premier He Lifeng, Bessent's usual counterpart, is a candidate to lead, with Ding Xuexiang, ranked seventh in the Chinese system, also mentioned.

The U.S. agenda, as described to Reuters, is narrow. Washington wants cooperation on monitoring cyberattacks directed by AI systems. It has floated a plan for U.S. and Chinese labs to police themselves and share information when those systems are used to break into networks. Export controls on advanced chips are being kept in a different room, according to people describing the design of the talks. The last dedicated U.S.-China AI dialogue was in Geneva in May 2024. Groundwork for a restart was laid when Trump and Xi met in May 2026.

Putting Bessent, not a science or defence official, at the head of an AI safety meeting is a tell. The Treasury secretary is the administration's main channel to Beijing on tariffs, investment screens and fentanyl precursors. If he also opens the AI file, safety language will travel with leverage language even if the briefers say the chip controls stay elsewhere. Beijing has already treated some American AI rules as industrial policy dressed as safety. That argument will be on the table whether or not anyone prints it in a communique.

Self-policing by labs is the weakest form of agreement two rivals can sign. It requires companies that compete for models, talent and military contracts to phone one another when a system starts writing intrusion code. It also requires both governments to trust that the phone call happens. The value of mid-September, if the meeting occurs, is not a treaty. It is whether the two sides can write down a working channel before Xi lands in Washington on the 24th.

Trump's second term has treated AI as both a commercial prize and a national-security risk. China has treated it as a sovereignty file. A meeting that discusses only cyberattacks, and not compute limits, is the overlap both sides can admit in public. The White House denial may mean the date is wrong, the lead official is wrong, or the administration does not want the session described before Xi's visit. All three are compatible with talks still being built.

Readers should hold the Reuters account and the White House sentence together. Planning is not a booking. A booking is not a result. What is new this weekend is that AI safety is being discussed as a possible agenda item in the two weeks before a Trump-Xi summit, with Bessent's name attached, and with a specific proposal about labs sharing attack data. That is the part that did not exist in last month's briefings.