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.

They propose to hoist a flag for one another: when a thinking machine in either country goes dangerously wrong, the other is to be told. I read the report twice, half in hope, half in irritation.

Sailors solved this problem long ago. A ship, whatever colours she flies, shows the yellow jack when fever is aboard, and no captain mistakes the signal for a trick, because the codebook is shared and the claim can be verified - send a boat, count the sick men. Two conditions: a vocabulary agreed in advance, and an inspection the warned party may perform.

Now trace the proposed mechanism. An incident occurs. Step one, detection: by what instrument, in whose service? Step two, classification: is this a threat to national security - and who declares it, when the declaration costs the declarer dearly? Step three, transmission. Step four, receipt: what may the receiving state verify, and by what boat? The proposal attends to the middle of the sequence and leaves both ends dark. No ministry hurries to publish its own machinery’s failures, and a warning that cannot be checked is indistinguishable from a feint.

Yet mark what is truly being attempted, which the officials themselves may not see. One cannot notify another of an incident unless both agree on what an incident is. The flag is trivial; the codebook is the revolution. Two rivals who define a common category accept, whether they intend it or not, a shared variable - and from shared variables one may write procedures that run on both machines alike.

So my hope is conditional, as all sound engineering is conditional. Give me the codebook - exact definitions, inspection rights, timing - and I will believe in the flag. Without them, it is cloth on a stick.

Two governments are discussing a courtesy: when their machines go wrong, they will tell each other. They call it a mechanism. A mechanism is a thing you build before the flood, not after.

Here is what I know about warnings. A warning is worth the hours between the sending and the arrival. Tell your neighbor the fire has started after the fire has crossed the road, and you have not sent a warning. You have sent an obituary.

And here is what I know about rivals. A signal line holds only when both ends want the same outcome. These two want different outcomes. So the arrangement will be tested not at the signing table but at three in the morning, when one side has an incident and silence serves it better. That is the moment to plan for. Not the handshake. The hour when the hand hesitates on the wire.

They are discussing while the machines are already running. The discussion is not caution. It is a way of feeling careful without being safe.

If they mean it, they will build for the lie. A second channel that does not depend on goodwill. Verification that does not depend on trust. A rule agreed now, in daylight, about what happens when the first notification turns out false - because there will be a first false notification, and the whole structure will stand or fall on what happens in the hour after it.

The stakes are not quarterly. If these machines fail badly, there is no road back, no night dark enough to move through unseen, no station on the far side. When the consequence is that final, the pace of the work must match it.

Build the wire. Test it in fair weather. Because the storm does not send notice, and neither, it turns out, do the men who promised to.