The US–China AI Incident Channel: What Was Actually Agreed

The United States and China have agreed to talk regularly about advanced AI and to set up a direct line for AI incidents. That is the whole of the commitment as the White House recorded it in a fact sheet issued late on Friday, September 25, after Xi Jinping’s state visit to Washington. The next meeting is due by November 2026.

Some observers have likened the channel to the Cold War hotline between Washington and Moscow. The comparison is premature. The fact sheet gives the channel a name and a purpose, but not a definition of an incident, an operator or a response time.

The White House fact sheet covers trade boards, fentanyl precursors, coal purchases and a pair of pandas for Atlanta. The AI material is two bullets near the end, under a heading about making the world safer. In our own words, they say three things:

Axios reported that Treasury Secretary Scott Bessent pushed for the channel in talks with Chinese officials the weekend before the summit. CNBC had earlier reported Bessent saying the US proposed an alert system for AI incidents. The fact sheet is the official record that both sides accepted it.

Only one row has a date. The rest are the details that make a hotline useful in practice. A channel with no agreed threshold can be used for everything or for nothing; one with no response time cannot be relied on during an incident, which is the only time it matters. Axios made the same point, noting that it is unclear which incidents would qualify or what either side would tell the other.

The fact sheet’s new term deserves a sentence of its own. “Super intelligence” is the White House’s wording, and Axios described it as President Trump’s preferred term. The Chinese side’s public framing, at least in the leader-level remarks, was different.

That reported remark talks about AI and its misuse, a narrower subject than the more capable future systems that “super intelligence” usually refers to. We have not seen a Chinese government text that uses the new term. If the two sides mean different things by the topic of the dialogue, the November meeting will have to settle that before it can settle what an incident is.

Those contract terms are where most businesses meet AI incident reporting today, and they vary widely. Our comparison of how fast 11 AI vendors must report an incident shows how uneven they are. For a company running its own agents, the more useful work is internal: knowing what you would count as an incident and what evidence you would keep, as set out in our guide to agent incident disclosure.

The government channel may matter later. If November produces a shared definition of a serious AI incident, that definition is likely to travel into regulation and procurement on both sides. A company that already records incidents in a consistent way will adapt faster than one that starts from nothing.

A two-bullet agreement attracts interpretation, and the interpretations tend to supply the parts the fact sheet leaves out. The illustrative readings below each add something the published record does not contain. They are patterns of commentary, not positions any named observer has taken.

The hotline that exists: the channel is described as a working line between the two capitals. The fact sheet says the two countries agreed to establish one, in the future tense, and names no operator, no launch date and no procedure, which is why the lede calls the Cold War comparison premature. The duty for companies: a compliance note tells AI teams they may now have to report incidents to a government channel. The channel is between governments, and nothing in the fact sheet says how, or whether, a company’s incident would reach it; the callout above gives the short answer, nothing yet. The shared definition: an analysis assumes both sides mean the same thing by an incident and by the topic of the dialogue. The fact sheet defines neither, the White House uses super intelligence and Xi’s reported remarks spoke of AI and its misuse, and we have seen no Chinese government text using the new term. The November deadline for rules: the next exchange, due by November 2026, is read as the date the threshold, operator and response time will be published. The fact sheet dates the meeting; it promises nothing about what the meeting publishes.

Each reading would be reasonable if the fact sheet said more. It does not, and the honest summary remains the one in the comparison table: one row with a date and five without one.

The readings also share a direction. Each one fills a gap in the record with the version of events that would make the channel more consequential than the text supports, which is the easy mistake with a short document about a large subject. Reading the fact sheet for what it says, rather than for what a hotline would need, is the discipline the rest of this post tries to keep.

The conclusion says to watch the November exchange for a definition and a response time, and the check is a short list drawn from the table’s open rows. Whether any text is published beyond a fact sheet, and if so whether both governments publish one. Whether an incident is defined, and at what threshold. Whether an operator is named on each side, and whether the channel is described as running rather than agreed. Whether a response time, or any obligation to respond, appears. Whether the two sides use the same term for the topic, since the super intelligence wording is Washington’s and the Chinese readout, as CNBC translated it, speaks of AI and misuse. And whether the next meeting after November is dated, which would show the dialogue has a cadence.

An illustrative outcome in which the two sides publish a joint statement naming operators and a definition of a serious AI incident would move most of the table’s rows from open to dated and would be the point at which the definition starts to travel into regulation and procurement, as the business section anticipates. An outcome in which only the next meeting is dated would leave the channel where the fact sheet left it, agreed and not running, and every open row in the comparison table would stay open.

In the meantime the work for a company is the work the business section describes and does not depend on the exchange: decide what you would count as an incident in the systems you run, what evidence you would keep, and what your AI vendors’ contracts already oblige them to tell you and when. A company that has done that will read whatever November publishes against definitions it already uses.

Put the November date in the calendar with the six checks beside it, and write your own incident definition before then. If the exchange supplies a threshold, you will be comparing it with something; if it does not, you will still have one.