MPs want some AI uses banned outright: what is in the 100-page report

The central proposal is a single, independent AI oversight body placed on a statutory basis. In the committee's view the current framework is fragmented and difficult to navigate, which leaves gaps in protection for people.

The second proposal ties duties to risk: the higher the risk of a system or model, the more demanding the obligations, applied across the whole AI lifecycle rather than only at launch.

The committee argues that some uses of AI are incompatible with human rights and should be prohibited entirely. The report names subliminal techniques and inappropriate use of profiling or biometric data as areas for action.

As examples of abuses that have already happened it cites the use of AI to create sexualised images of women and girls, and the scanning of people's faces without their consent.

Committee chair Alex Sobel, a Labour MP, framed the complaint broadly: nowhere in the world, including the UK, has a legislative and regulatory approach to AI that is fit for purpose. The technology moves with such speed and complexity that its impact is hard to predict, and the country is unprepared for the consequences.

Labour's 2024 general election manifesto promised binding regulation for the handful of companies developing the most powerful models. The July 2024 King's Speech committed the government to appropriate legislation, but no dedicated AI bill appeared in this year's King's Speech.

Education minister Georgia Gould responds that the government has been regulating: action against non-consensual sexualised deepfakes, a bill linked to cyber security, and the AI Security Institute testing models as they come into the UK and working through future challenges.

The report lands while the alarm is also being raised inside the companies. Jacob Coxon, an AI researcher who quit Anthropic, told the BBC that staff developing these systems were genuinely frightened for the future of humanity. An Anthropic spokesperson said the company builds models with some of the strongest safeguards in the industry.

A day earlier Anthropic's chief executive Dario Amodei proposed his own plan: global regulation, industry-wide rules, independent monitoring of models as they are developed, and a slower pace. Sam Altman of OpenAI and Elon Musk both said they agreed.

One long-running concern reappears in the report: a model trained on large datasets scraped from the internet inherits the bias in that material and can reproduce racist and sexist patterns, as well as help create and spread misinformation.

A British law, if it arrives, will not apply in Russia. But regulators are moving in the same direction everywhere, and preparing early costs almost nothing.

• Sort your use cases by risk. Interface hints and draft copy are one thing; automated decisions about people such as hiring, credit or access are another, and the second kind needs a human in the loop.
• Record what data the model runs on and where it came from. Provenance is the first question both regulators and clients ask.
• Treat biometrics and profiling as the last resort and only with explicit consent: these are the uses that appear on prohibition lists.
• Keep a decision log: model version, prompt, output. Without it you can neither investigate an incident nor show that you acted in good faith.
• Name a person responsible for the system after launch. Lifecycle oversight is mostly about operation, not about the release.

The committee does not pass laws, it states a demand. The decision sits with a government that so far prefers targeted measures to a general act. Companies are better off not waiting: the steps above pay for themselves regardless of which text is eventually adopted.