The Kill Switch Nobody Was Ever Going to Pull
Somewhere inside the Cabinet Office there is a position, unpublished in full but summarised often enough, explaining why the government will not build a kill switch for artificial intelligence. The phrase is doing work. It makes the idea sound like something from a Bond film, a red button under glass, rather than what it actually means: a legal power to compel a company to stop a system doing something demonstrably dangerous. Reject the button and you have also rejected the power. Nobody voted for that trade, and nobody will be voted out over it.
The occasion was a joint committee of MPs and peers, sitting as Parliament's human rights watchdog, which concluded that existing law is not fit to manage what AI can now do to people who never agreed to be governed by it. That is a serious finding from a serious body. The government's answer — more work is under way, frameworks are evolving, international partners are engaged — is the answer it gives to almost everything. Translated: the current arrangement suits the people currently running it, and asking who is accountable for it is considered bad manners.
It is worth being precise about what is at stake, because "AI safety" has become vague enough to cover both a chatbot inventing a legal precedent and a surveillance system sorting a population by risk score. The US State Department's own risk profile on AI and human rights lists the plain uses: arbitrary mass surveillance, censorship of the information environment, discrimination entrenched in criminal justice, immigration, welfare and finance decisions made by systems nobody outside the vendor fully understands. The United Nations adds a sharper one. AI can manipulate voters and shape opinion during the very elections meant to be the citizen's one working lever on all of this. The lever is itself now a target.
None of that is speculative. It is the working consensus of the bodies that study the technology for a living, from the Freedom Online Coalition's joint statement warning that AI without safeguards threatens freedom of expression, assembly and religion, to researchers cataloguing biometric surveillance creeping into asylum processing under national-security exemptions written to survive scrutiny rather than invite it. The harms are documented, the remedies are discussed, and the power to act sits somewhere just out of Parliament's reach.
The government's defenders have a case, and it is not a stupid one. Build in a hard off-switch, they argue, and you build in a competitive disadvantage against states and firms under no such obligation. You also risk freezing a technology whose beneficial uses — in diagnostics, in fraud detection, in flagging where a conflict is likely to break out before it does — are real and are being lost in the panic about the harmful ones. Better, on this view, to work through the frameworks we already have, imperfect as they are, and improve them in increments alongside industry and civil society, rather than legislate in haste against a target that will have moved twice before the ink dries.
The trouble with that case is that it was made, almost word for word, about the last general-purpose technology British governments were frightened of controlling: the atom. In 1954 a country still rationing sugar decided nuclear power was too dangerous to be run on ministerial discretion and goodwill, and put the whole business on a statutory footing. An Atomic Energy Authority with legal licensing powers, inspectors who could order a reactor shut down, a chain of accountability running back to a minister who answered to the Commons. Nobody argued this would kill Britain's nuclear industry. It did not. What it did was make clear whose signature was on the licence when something went wrong.
Artificial intelligence, unlike a reactor, has no containment vessel and no fence around it, which is precisely why the analogy holds rather than fails: the risks are more diffuse, so the case for statutory power over them is stronger. Instead Britain has a Safety Institute inside the Cabinet Office with no statutory footing at all, answerable to a minister who can reshuffle its remit by press release, studying harms that a parliamentary committee has already said the current law cannot reach. That is not caution. It is the 1954 decision taken in reverse, dressed up as prudence.
Ask who decided that AI safety should live in the Cabinet Office rather than in statute, and the trail runs through several written answers and a great deal of "further work is needed," and arrives at no desk in particular. A civil servant cannot be sacked at a general election. A minister who has delegated the judgment to a non-statutory institute cannot meaningfully be blamed for what it fails to prevent.
The voter is left holding the one power democracy is supposed to guarantee her against exactly this kind of harm, and finds it does not reach the room where the harm is being managed. The government's line is that the technology moves too fast for Parliament to legislate safely. It might try, once, and see what it can hit.