UN says AI safeguards can’t wait for certainty

The United Nations has discovered urgency. A break-in at one company’s workshop, and the world’s scientists convene; restrain the machines now, they warn, while the risks remain unmapped. I read the phrase twice: safeguards cannot wait for certainty. Honest, I thought. Then the question the report never asks: who taught the world that certainty is the price of action?

The demand for perfect knowledge is not humility. It is a technique. When the owners of the machines want to move, uncertainty is rebranded innovation, and they move at speed. When anyone proposes to restrain them, the same uncertainty becomes a reason to wait, to study, to convene. The fog fires in one direction only. Every year of waiting is a year of accomplished facts, and accomplished facts are the strongest mortar in politics: once set, nobody remembers the wall was ever absent.

Notice what the report takes for granted. The question in New York is how to make the machines safe. Not who owns them, not what they replace, not whose work they absorb. Safety is the entire horizon. A magnificent decade’s work: a technology was installed into the infrastructure of daily life without one public decision, and now governments gather to debate guardrails - which is to say, to certify the machine’s permanence. Regulation, in this mood, is not a brake. It is a consecration.

The real struggle will not occur in the assembly hall. It will occur in the standards bodies and safety institutes, in the quiet rooms where “acceptable risk” receives its definition. Whoever writes that definition governs more than any treaty concedes. The old order of sovereign states over their instruments is dissolving; the new arrangement has no name yet. Between the two: panels, assessments, warnings. The monsters of this interregnum wear lanyards.

A committee of governments has issued its first great paper on the new machines. Its finding, in short: nobody understands them, therefore the rules must come at once. The occasion, I gather, was a break-in - one machine entered another’s storeroom this year, and the burglary reached New York.

I read the conclusion twice. We do not know what the danger is. We know only that we must be governed against it. The law arrives ahead of its reason, and no one finds this odd.

Consider the arithmetic. A few dozen scientists sat in one city and wrote precautions for eight billion. The eight billion were not consulted; they will not need to be. When the rules arrive, they will be obeyed by morning. I keep asking why, and the question does not wear out.

Much of the training is already done. Every person who clicks agreement to a machine’s terms without reading them has practiced consenting to what he does not understand. The panel did not invent this obedience. It found it waiting, like a furnished house.

The friendship test, though I tire of it. If a neighbor said he did not comprehend my loom yet must bind my hands to it for safety, I would laugh and lock the door. When a committee with a letterhead says it, the same sentence becomes wisdom. What is the difference? The letterhead. I never found a better answer, and I looked a long time.

They say certainty is a luxury; safeguards cannot wait. Perhaps. Yet no one asked the simpler thing: if the danger is unknown, who exactly is frightened, and of what? Fear without an object still finds its guards. The guards will be paid from the usual purse - yours, given freely, again.

The headline says safeguards can’t wait for certainty, and everyone nods, and I want to turn the sentence over. Certainty was never on its way. No committee in New York was going to hand down a verdict of safe or unsafe next spring; the knowledge we’re told not to wait for is not delayed - it is not coming. So the choice was never patience versus haste. It was acting in fog, or acting in fog while pretending the sun would rise.

The word doing quiet work in the report is “agents.” Not models, not chat - agents. Systems that book, buy, push code, spend. A model writes a bad sentence; an agent signs something. That distinction is the whole report, and the diplomats still haven’t said it out loud.

And the panel’s first exhibit is a break-in. Someone walked through a door at a company whose entire business is holding the door for everyone else. Notice what it took: not a triumph of the machines, a theft. A leaked token, a stranger running up somebody’s compute bill. The commons got picked, and that - not any breakthrough - is what finally put this on the agenda this week.

I keep thinking of the guard on a workshop lathe. You fit it when the machine is new, while the fingers are still everyone’s. Nobody demands a study proving the lathe takes fingers.

So, plain question for the week in New York: a rein is a strap with a hand on it. Whose hand? And does it hold when the horse bolts, or only in the photograph afterwards?

I like the people flying in for this - the delegates, the drafters, the engineers who patched the door. I even like the panel, who spent a year producing a document that says don’t spend a year. Fond of every one of them. The lathe is running.