An AI incident line between the U.S. and China

Treasury Secretary Scott Bessent said Monday, September 21, that the United States and China had agreed to establish a formal AI dialogue, including an incident line for communicating about AI safety problems. He said officials would meet again in roughly two months in Shenzhen to discuss dangers and response protocols, according to Reuters’ report of his CNBC interview.

That advances the account given on Sunday evening. After meeting Chinese Vice Premier He Lifeng in New York, Bessent described a U.S. proposal for notifications about AI incidents affecting national security. The Associated Press reported those discussions ahead of the planned September 24 Trump–Xi meeting.

An agreement to talk is worth following. It is also too early to describe this as an operational emergency system: the reporting reviewed for this edition does not establish its launch date, notification thresholds, or public reporting rules. Monday’s account is Bessent’s description of the agreement, rather than a published bilateral protocol.

The practical questions are straightforward. Who must report an incident, and how quickly? What information crosses the line? Who can act on a warning? And can researchers or affected people learn enough afterward to check whether the response worked? Those are our tests for the next announcements, not features already promised by both governments.

Meanwhile, a separate oversight proposal has acquired a named partner. In a September 18 announcement, Anthropic said Accenture’s Faculty business would lead embedded evaluations, including red-teaming, alignment assessments and safeguard testing. Anthropic says it will pay Accenture directly. It also acknowledges that access standards, reporting arrangements and the wider funding system remain unsettled, and says it will continue training and releasing frontier models.

That combination makes the arrangement useful to scrutinize. Inside access could help evaluators see things that a limited external test misses. Direct payment by the company being examined also makes the protections for independent findings especially important. Neither observation proves that the arrangement will succeed or fail; published operating rules and actual findings will be more informative than the partnership announcement alone.

For ChatGPT users, these developments do not change today’s settings or establish that any model is safe. They are concrete places to watch whether oversight becomes inspectable and actionable. Our Safety & Society guide separates everyday precautions, documented harms, policy proposals and frontier evaluations so you can follow the evidence behind those claims.