On: US and China Discuss Alerting Each Other to AI National Security Threats

Two governments sat down to discuss telling each other when their machines do something dangerous. Discussed. Not built - discussed. In the lab, “we talked about building the detector” detects nothing.

Here’s the thing you have to understand about warning systems: before you can notify, you must detect. Before you can detect, you must know what you’re looking for. So - what is an AI incident? Point to it. A missile leaves a streak across the sky. A seismometer can tell a blast from an earthquake because the ground wobbles differently. What does an AI accident look like? No flash. No wobble. Possibly nothing at all - just a machine quietly doing the wrong thing very well, somewhere nobody thought to watch.

Wait - that’s not quite right. Maybe the incident is visible to the country it happens in. Fine. Then ask the second question: why would they tell us? A notification says, in effect, “our machine misbehaved in a way that might hurt you.” That is a confession of weakness handed to a competitor. You send it only when silence is more dangerous than the confession. Which means the channel carries exactly the incidents that were never going to matter, and stays silent for the ones that would. A bell that rings only when nothing is wrong.

I would ask the officials for one sentence: define the incident. If they can’t define it, they can’t detect it. If they can’t detect it, the mechanism is a communiqué wearing a machine’s clothes.

Still, talking is not nothing. Two rivals agreeing that there is something worth watching for - that’s the first datum, and I’ll take it.

But a warning system you cannot test is a hope, not an instrument. Build the detector before you promise the bell.