Britain Is Building an AI Regulator Without a Regulator: Three Lords Amendments, One Missing Body

Britain has never passed an artificial intelligence law. As of October 2026, it governs frontier AI through voluntary agreements, published codes of practice, and the expectation that the companies building the most powerful models behave responsibly. Three members of the House of Lords are now dismantling that arrangement with amendments to a cyber security bill.

In the six days between 9 and 13 October, three peers from three different benches tabled three amendments to the Cyber Security and Resilience (Network and Information Systems) Bill, and each one hands a job to the AI Security Institute that no British statute has ever given it. Taken together, they amount to the architecture of a frontier AI regulator, assembled in instalments during the report stage of legislation that was supposed to be about hospitals and water companies.

Three amendments, three rival designs

The first came from Lord Tarassenko, a crossbench peer, tabled on or before 9 October, and it is procedural. Before a frontier AI product could enter a regulated essential service in the United Kingdom, the developer would hand the Institute its evaluation results, network monitoring records, and evidence of a tested means of suspending the product. The Institute would get 28 days and a duty to recommend a block where it finds a serious unmitigated risk. In effect, a pre-deployment audit with a statutory veto.

The second arrived on 12 October from Baroness Kidron, also a crossbencher, and it is substantive: it lists nine capabilities and says no person may make available in the United Kingdom an AI product or service that has any of them, including autonomously harming an essential service system, evading human oversight, supporting terrorists or hostile actors, or autonomously self-replicating in a way that could gain the ability to cause such harm. Two subsections later comes the compliance route: the developer demonstrates safety training against the listed behaviours, a tested suspension mechanism, and human-control-first design, and reports it all to the Institute before launch. On the Institute's advice, the Secretary of State could switch a non-compliant product off.

The third, tabled on 13 October by Lord Clement-Jones is the most complete. His 31-line clause would require the Secretary of State to ensure the Institute has power to set and keep under review safety standards for certain frontier AI systems, to inspect those systems and the practices of the people developing or deploying them, and to require remedial action, including suspending or restricting provision, where they fall short. Standard-setting, inspection, enforcement: that is the full set every functioning product regulator has.

Three peers have shown that statutory AI regulation fits inside a cyber security bill in a page each. The only reason the Institute is not a regulator by the end of the month is that the House has not voted yet.

The body these powers are handed to does not exist in law

Here is the defect running through all three. The AI Security Institute is a directorate of government. It publishes research and evaluates models under voluntary agreements, but it has no statutory personality: it cannot be sued in its own name, holds no statutory duties, and no appeal route runs against its views. None of the three amendments creates such a body. Lord Clement-Jones's is the only one that gestures at due process, requiring regulations on notice, representations, and appeal, but regulations cannot confer legal personality on a directorate.

This is not a theoretical objection. In September 2026 the Joint Committee on Human Rights concluded that developers engage with the Institute voluntarily, that it holds no statutory power, and that no regulator can test a system before release or prevent a release judged to carry unacceptable risk. That arrangement held until it did not: on 1 September 2026 a frontier model launched without the Institute being given pre-release access, the first time the British testing body had been left out. Every one of these amendments is written in the shadow of that date.

They overlap, and that is a problem

Read together, a second defect appears: they are not complementary. Lord Tarassenko's would gate deployment on an audit against a serious unmitigated risk test. Baroness Kidron's would gate market access on a self-written safety case against nine capability red lines. Lord Clement-Jones's would gate continued operation on conformity with Institute-written standards. Passed together, a frontier developer would face three overlapping regimes, administered by the same non-statutory body, against three different tests. The House will have to choose, in a single day: report stage is scheduled for 26 October 2026, according to the UK Parliament Bills API.

The counting matters. As of 13 October, 47 amendments stand on the paper in eight names across four benches, and the government has tabled none of them. That argument is harder to make now that three peers have shown the powers fit inside this bill, at a page each.

A cross-party bloc is forming in the background

The signatures tell their own story. On the 12 October list, Lord Clement-Jones appeared on five amendments, all his own. On the 13 October list he appears on twelve, having added his name to six clauses led by three other peers, including Baroness Kidron's red lines clause and Lord Tarassenko's audit clause, plus incident reporting amendments led by Conservative Baroness Harding of Winscombe. Two days earlier, Kidron had made the same move in reverse, adding her name to 21 of Harding's incident reporting amendments. In three days the paper has acquired an unannounced cross-party bloc: a Conservative backbencher leading on incident reporting, two crossbenchers on capability red lines, and a Liberal Democrat frontbencher signing nearly all of it.

The government, for its part, issued a written ministerial statement on 12 October saying it is considering how the regime can be strengthened so that AI cyber and loss of control risks are better understood and protected against, naming three examples: near-miss reporting, reducing AI-driven risk across organisations, and sharing lessons from incidents. The same day, the Minister of State for AI told AI companies at Imperial College London that if they cannot adequately control the risks, they should not release the system. Baroness Kidron's amendment is that sentence turned into law, and the government has not signed it.

What happens next

Everything now converges on 26 October. At report stage in the Lords, the House will decide whether Britain's first statutory AI powers live inside a cyber security bill about essential services. If none passes, the voluntary regime survives, with a public record showing Parliament declined ready-made alternatives. If one does, the United Kingdom gets a legal basis for inspecting frontier AI systems and ordering them suspended, administered by a body the law would still need to create.

That last point is the one to watch. Whatever passes on 26 October, giving a government directorate the powers of a regulator without giving it the legal existence of one is a fixable defect, and fixing it is the first thing the Commons will have to do when the bill arrives there. Britain is about to regulate frontier AI. The only question left is whether it builds the regulator first or pretends the directorate is one.