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

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.