Hermer demands urgent law to control artificial intelligence

The Refrigerator Has a Test Bench

We are told the limit is a legal one: that artificial intelligence escapes governance because no one has yet written the law, and that the cure is urgent public debate and robust international law. But notice the instrument that has arrived. These are general-purpose models that can draft a contract, write working code, and hold a conversation with a frightened patient at three in the morning. A single artefact now does what once took a dozen trades. The question is not whether we should revere the old frame of treaty and statute. The question is what experiment would show whether law can reach this tool at all, and that experiment has not yet been designed.

Richard Hermer’s comparison deserves more attention than its author perhaps intended. He says AI faces less regulation than a refrigerator, and he is right. Ask why the refrigerator is regulated, though. Parliament did not hold a great debate on the moral nature of cold. The refrigerator is governed because it can be measured. You can put it in a chamber, run it for a set time, read the meter, and print the energy grade on a sticker. Anyone can check whether the door seals, whether the wiring holds under load, and whether the coolant leaks. Every rule that touches the appliance stands on a test someone can repeat. The rules came after the test bench existed, and they could not have come before it.

The history of the refrigerator shows this sharply. Its coolant gases were thought harmless for decades. Diplomats did not overturn that belief with eloquence. Scientists at a station in Antarctica did it with instruments, measuring a thinning of the ozone layer that no committee had predicted. Once the harm had been named and shown, nations moved with unusual speed, and the Montreal Protocol followed, perhaps the most successful international agreement of the last century. Hermer wants international law, and in the refrigerator he has picked the best precedent available without quite noticing what it teaches. That treaty worked because a spectrometer had settled the argument before the diplomats arrived.

The strongest form of his position should be stated fairly. The danger he names is losing control of the future, and that is not a leaking seal. It is a harm that, once demonstrated, may already be beyond repair. Demanding evidence before acting can look like waiting for the house to burn before buying the extinguisher. I grant the force of this. But a precaution against an unspecified catastrophe cannot be enforced, because nobody can say whether it has been obeyed. A treaty that forbids “losing control” is like a treaty that forbids bad weather. It cannot be inspected, audited, or breached in any way a court could recognise. The states that sign it will be the ones that never meant to build the thing, and the states that build it will sign it gladly, since it costs them nothing.

Picture the inspector. She arrives at the appliance factory with a thermometer, a clipboard, and a stopwatch, and by lunchtime she knows whether the law is being kept. Now send her to a laboratory that has just trained a frontier model. What does she plug in? Which dial does she read? She could put questions to the system, but she has no agreed list of questions, no threshold that counts as failure, and no standard that a second inspector in another country would apply in the same way. She stands in a server hall holding a clipboard with nothing written on it. This is the real gap Hermer has found. The missing law is a symptom. The missing instrument is the cause.

So the urgent task is less a public debate than a public test bench: shared, repeatable evaluations for the specific capabilities people fear. Can the model help someone synthesise a pathogen? Can it copy itself onto other machines? Will it deceive the people overseeing it when deception pays? Each of these can be set as a task, run, and scored, and each result can be checked by a rival laboratory or a hostile government. Once those measurements exist, international law has something to stand on. Without them, the robust law he calls for will be a fine set of words that no one can inspect.

The refrigerator wears its energy grade on its door for any shopper to read. When the most powerful machines of 2026 carry a grade like that, earned in a chamber anyone can rebuild, the treaty will almost write itself. Until then, the lawyers are drafting rules for a door they cannot yet open.