What an AI incident channel needs before it counts as one

Governments are starting to talk about notifying each other when an AI system does something dangerous. The word that gets used is "hotline", and it carries an image: a line that rings, someone who picks up, a decision that follows. Almost none of the proposals on the table describe that. They describe an intention to describe it later.

The current example is concrete. Ahead of the 24 September 2026 Trump–Xi meeting, the United States put forward a mechanism for notifying China about AI incidents. Beijing confirmed through Xinhua that AI was discussed. It did not say the proposal had been accepted. FuturPulse worked through what Trump was actually being asked to approve before the Washington summit, and the honest summary is that the published record establishes a topic, not a channel.

That is not a criticism of the idea. It is a description of how far along it is — and the distance is worth making explicit, because anyone building incident response inside a company already knows what the remaining work looks like.

The four things any incident channel has to specify

If you have ever run an on-call rotation, this list will be familiar. It is the same list at the level of states.

A trigger. What counts as an incident? A model that exceeds a capability threshold, an autonomous system that acts outside its authorised scope, a deployment that produces physical harm? Until the trigger is written down, every party gets to decide privately that nothing has happened yet. A channel with no defined trigger cannot fire, and — more importantly — cannot visibly fail to fire, which means nobody can be held to it.

A named endpoint. Who receives the message, at what address, with what authority to act on it? "The two governments" is not an endpoint. Real channels name a desk and a deputy, because the person who answers at 03:00 is rarely the person who negotiated the agreement.

A message format. What has to be in the notification? A timestamp, a system identifier, an observed behaviour, a confidence level, a recommended action. Without a format, the first real notification arrives as prose, is read as a diplomatic signal rather than a technical report, and is answered in kind.

A response deadline. How long may the other side take to acknowledge? An unbounded acknowledgement window converts an incident channel into a suggestion box.

No public document sets out any of the four for the US–China proposal. That is the measure of where it stands.

Why the gap keeps being missed in coverage

Because the label does the work. "Hotline" is a familiar object, so readers fill in the operational detail from the word rather than from the document, and the story gets covered as though a mechanism now exists. The same move happens with "task force", "czar" and "framework". The correction is mechanical: before treating an announcement as an instrument, ask which of the four items above appear in a published text, and note the ones that do not.

The same test applies to your own vendors

If you run AI systems in production and depend on a provider, you have an incident channel with them too — and it is worth checking against the same four points. Most status pages define no trigger for model-behaviour incidents, name no endpoint beyond a support queue, specify no format, and promise no deadline. The state-level version of this problem is the one in the headlines; the vendor-level one is the one that will actually page you.

FuturPulse covers AI policy with that separation in mind: what a published document establishes, and what it merely announces. Ongoing coverage is at FuturPulse.