An AI Emergency Hotline Between Washington and Beijing: Could It Prevent a Crisis?

1 What the reporting actually describes

Axios reported that the United States has proposed a US-China AI notification mechanism, in effect a hotline for artificial intelligence incidents, and that Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng discussed it in New York on September 20 and 21, 2026. Trump science adviser Michael Kratsios put the aim as avoiding shared risks. This is a reported proposal, not an agreement with published implementing text.

2 The uses that would make it worth having

The case for a line rests on incidents that cross borders and resist attribution: dangerous model behaviour, runaway autonomous agents, cross-border cyber operations and bioweapons development. Each shares one property: the harm arrives before anyone agrees who caused it. A notification mechanism is a tool for that gap.

3 The trigger is the missing part

A mechanism only works if something sets it off, and exactly what would trigger a notification has not been publicly revealed. Without published triggers, two governments cannot rehearse the same response, and every incident becomes a separate argument about whether it qualified. WIRED reported on September 23 that the administration has internally discussed criteria for obligations and enforcement, partly so the arrangement does not become one where only Washington alerts Beijing.

4 Beijing has not committed, and the clock is short

China's response so far is noncommittal: Xinhua acknowledged that AI was discussed but did not say Beijing accepted the proposal. Officials cited by WIRED expect several weeks to finalize a framework, aiming for something functional by the next Bessent-Lifeng meeting in Shenzhen in two months. Between proposal and working channel sit the deciding questions: who staffs it, and what a notification obliges either side to do.

5 What a usable design would require

The Carnegie Endowment's recommendation is specific: a text-based channel rather than a voice line, tiered so routine questions do not require a leader-level call, continuously staffed, and organised around defined triggers. Text leaves a record, which makes verification possible. Tiers stop small events from escalating by default. Permanent staffing removes the question of who is awake.

6 The record of failure is the real test

Carnegie's Lucy Luo argued on September 18, 2026 that US-China hotlines have repeatedly failed at the moments they existed for. After the 1999 bombing of the Chinese embassy in Belgrade, US officials could not reach counterparts; after the 2001 Hainan Island incident, calls went unanswered for 12 hours; in 2022 China suspended the military hotline after Pelosi visited Taiwan.

7 The open question

Whether this channel is useful turns on verification and trust, not on whether a line exists. Two tests settle it. Are the trigger conditions public enough that both governments know in advance what must be reported? And does a notification create an obligation to respond, or only to transmit? Until the triggers are published and Beijing commits rather than acknowledges, the proposal describes a channel, not what either side would do on the first night it is needed.