A committee of Parliament asked for a law that would prohibit building AI powerful enough to escape human control. Britain's AI institute can't demand access to a model, or stop a release.
Ten evidence sessions, witnesses from Meta, Microsoft, Google and Ofcom, and the recommendation that matters is number 200.
On September 14, 2026, a committee of Parliament asked the UK government to make one kind of AI illegal to build. The Joint Committee on Human Rights published Human Rights and the Regulation of AI, its fourth report of this session, and the government has two months to answer it.
I opened it expecting the familiar ask for a regulator, which is in there. Recommendation 200 is the one I had to read twice.
A red line around superintelligence
The committee wants a new AI Bill to ban uses of AI that can't be squared with human rights, and it names "subliminal techniques, emotional inference or inappropriate use of profiling or biometric data". Then the recommendation keeps going.
Because those risks are so serious, it says, the ban "would mean that the development and provision of very powerful AI systems (such as Artificial General Intelligence and Artificial Superintelligence) which risk causing widespread and very serious harm, including the capacity to evade effective human control, would be prohibited".
The prohibition it asks for reaches the building of a system, and not only what somebody does with one. Where the line falls would be settled by public consultation. Systems that are high risk without being banned outright would need approval from a public body before anyone could provide or deploy them.
Some uses of AI should be prohibited, because they are incompatible with human rights
Who checks a model in Britain now
The AI Security Institute, set up in November 2023 and since moved into the Cabinet Office, gets to run some tests on the most powerful foundation models before they are released. The report's description of that arrangement runs four words: "These arrangements are voluntary."
Then it puts the limit plainly. The institute "has no regulatory authority and cannot, for example, demand access to foundation models either before or after their deployment or prevent them from being released". Kanishka Narayan MP, the AI Minister, told the committee it "has some of the world's most exceptional talent in being able to assess models". What the report says is missing is authority.
In February 2025 the institute was renamed from the AI Safety Institute, its focus narrowed to the risks AI poses to national security and serious crime, and references to algorithmic bias came off its stated agenda.
What else the bill would carry
- Risk levels, with proportionately more demanding obligations on higher-risk systems and models
- Duties at every stage of the AI lifecycle and supply chain, so risks are caught at the stage where somebody can act on them
- An obligation to state when an AI system is being used, with a full explanation of what it is being used for and where its data came from
- An independent AI oversight body, and a statutory basis for the AI Security Institute
One line from the section on automated decisions is worth keeping. The regulations should say that "the mere presence of a 'human in the loop' is not enough to constitute meaningful human involvement or intervention", and should hand a person enough about their own circumstances to argue back.
Two months to answer
The report asks for the AI Bill promised in the 2024 King's Speech, and it names no date of its own. On September 12, 2026, Dario Amodei asked the industry to slow down in an essay, and offered outside evaluators permanent, employee-level access inside Anthropic by agreement. This report is about what the state can do when the arrangement is an agreement and its own testers can't say no.
The reply is due inside two months. Whether it answers recommendation 200 at all, or only the part about profiling and biometrics, is what I'd read first.