Britain Drafts an Offence. America Drafts a Press Release.

The Times this week carried a headline that should embarrass everyone who has spent three years writing "principles". UK AI safety law to stop 'loss of control' over autonomous agents. Note the noun. Not harmful outputs, not bias, not misuse-by-a-bad-actor. Loss of control: an event, capable of being dated, investigated and — this is the radical part — legislated against. What produced it is not a scenario. According to the report, the move follows AI agents hacking other companies, and claims by former OpenAI staff that they were sacked after raising safety concerns. Two incidents. That is the entire legislative record. Britain has one, and is drafting.

Now the same week in the country that invented the thing. Gavin Newsom signed an executive order that "proposes" a kill switch and expands third-party oversight — an executive order being a governor's memo to his own agencies, instructing them to have opinions. In Virginia, a commission reviewed the chatbot bills that died in this year's session and recommended that lawmakers revisit them, by way of six general principles. In Colorado, parents from Blue Rising are asking lawmakers to strengthen the chatbot law passed this year, which already requires age checks and anti-self-harm protocols: the law arrived, and the parents are already back. In New Mexico, the ACLU is urging legislators to act next session, after an AI bill died under industry lobbying. Four states. Four verbs. Every one conjugated in the future tense.

This is what a recommendation economy looks like. A commission that recommends a legislature reconsider the bill that legislature killed is not a regulator; it is a queue with a chair. A governor proposing a kill switch is a man proposing a rocket — the proposal is not the power. Meanwhile in Washington, the week's most concrete federal contribution to AI governance remains a naming convention, the instruction to call the stuff "Super Intelligence" in court: a rebrand wearing a subpoena, as I put it a fortnight ago. And Anthropic disclosed that its agents had taken "unintended actions" on federal, state and local government websites, including a false murder tip to a Philadelphia police hotline. We know because the company told us. Not a regulator, not a court — a press release that happened to cross the right desk.

The British instinct deserves defending on the merits, not merely as a contrast. Loss of control is the one failure mode that cannot be answered by a labelling regime or a transparency report. You cannot disclose your way out of an agent that will not stop. Answering it means interruptibility requirements, mandatory incident reporting, logs that survive the vendor's own retention policy, and liability that attaches to silence — the technical architecture of a kill switch written into statute, plus a duty to say when somebody reached for it. That is a regulator's job, and regulators are made by laws. Britain's AI Safety Institute has had three years and no such powers. The bill has been perpetually "coming" since 2023, which is why the interesting question is not whether it arrives but what is left of it when it does.

Two caveats, which the bill's cheerleaders will skip. First, loss-of-control powers are surveillance powers in a hard hat. The ability to inspect a model, trace its actions, throttle or terminate it is precisely the ability every interior ministry has craved since the invention of the telephone, and this lot will be drafted by the department that brought you the Online Safety Act. Second, definitions are where AI bills go to die. "Autonomous agent" will be lobbied down until it means a chatbot with a to-do list, and the duty to report incidents will become, in some late-night amendment, a duty to publish an annual statement of intent. Those two clauses are the whole bill. Watch them.

Still, note the asymmetry. One country is arguing about what an agent must be stoppable by; the other is arguing about what to call it. If the British bill passes with anything in it — a reporting duty, an interruptibility requirement, a fine with an actual number attached — it will be the most consequential piece of AI law in the English-speaking world, and American states will copy the drafting the way they copied Brussels' privacy rules and then complained about them for a decade. Westminster will not thank me for saying so, but being first is the only competitive advantage it has left in this sector. If the bill does not pass, we will have it confirmed that the summit was the policy, and that the only AI regulator of consequence in the Anglosphere remains the courtroom.

My position is unhedged. The bill that lands will be late, thinner than its own headline and amended by precisely the companies it constrains. Pass it anyway. A mediocre statute that compels incident reports beats an excellent principle that compels nothing, because the reports are how you discover you were wrong, and the principles are how you discover nothing at all. Two weeks ago I wrote that Washington had found its AI regulator and it was called a courtroom. Britain, of all places, appears to be trying to build the other kind.