UK lawmakers want a regulator that could pause AI training and pull risky models off the market
Britain's parliament is pushing the government toward a legal power it does not have today: stopping an AI company mid-development. For OpenAI, Anthropic, Google DeepMind and Meta, that would add a new kind of launch risk in a major market.
The House of Commons Business, Innovation, Science and Trade Committee published its report, Guardrails for Growth: Making AI Safe, on November 18. It asks ministers to introduce an AI Safety and Security Bill once the next parliamentary session opens. The committee's chair, MP Liam Byrne, said the UK cannot give AI leaders "unfettered freedom to behave like modern-day Dr Frankensteins."
The report is not binding. The government has not committed to the bill, and its formal response is still outstanding. But the pressure built the next day, when MPs held what Carmarthenshire News Online described as the first debate on regulating advanced AI in the Commons chamber.
What the committee wants a regulator to do
The core proposal is a new statutory regulator, separate from the AI Security Institute (AISI), the government body that tests frontier models, the most advanced AI systems. AISI would keep supplying technical evidence. The new body would enforce.
Its proposed powers are broad. It could require independent checks on a developer's safety claims. It could delay, restrict or prohibit high-risk training, testing or deployment in the UK. And it could suspend services whose risks cannot be contained.
Developers of systems designated high-risk would have to back up their safety claims with evidence and give outside evaluators real access to their models. Any material change to a model's abilities or safeguards would trigger a fresh review. Oversight would start during training and internal use, not just at public release.
The committee also wants mandatory reporting of AI incidents and near misses, and protected channels for staff to raise concerns with the regulator. Former industry employees told MPs that restrictive contracts and financial pressure discourage them from speaking out.
The report lists ten gaps in current UK protections. They include no legal definition of unacceptable harm, no power to demand access to models, weak liability rules and no power to halt unsafe systems.
Why it matters to AI companies and their customers
Today, Britain's oversight of frontier AI runs on goodwill. AISI has tested some of the most powerful models, but only because companies agree to it. It has no authority to stop a release. The committee found that companies apply different safety thresholds and judge for themselves whether their own systems meet them.
Representatives of OpenAI, Google DeepMind, Anthropic and Meta gave evidence to the inquiry. Byrne wrote that competition between labs, including ahead of stock market listings, makes it hard for any single firm to slow down on its own.
If the proposals became law, the business effects would be concrete. A model launch in the UK could be delayed pending review. A service already in use by British companies could, in principle, be suspended. Firms building products on top of frontier models would carry some of that risk, not just the labs.
The report also tackles models that companies release for anyone to download and run, which cannot be recalled once copied. It suggests the government should be able to act against specific hosts and services, and to criminalise harmful misuse.
The committee frames all of this as pro-growth. It cites OECD productivity estimates suggesting AI could make the UK economy up to £400 billion larger after a decade, and argues that serious failures would erode the trust that adoption depends on. Byrne also argued that Britain could build an export business in AI testing and assurance.
What comes next, at home and abroad
The committee wants ministers to publish, within three months, what the UK will seek from its G20 presidency, which begins next month. It wants an international convention on AI safety considered, but says talks abroad must not delay rules at home. Prime Minister Andy Burnham has already proposed a "Manchester Compact on AI" at the 2027 G20 leaders' summit.
At the Asia-Pacific Economic Cooperation summit this week, President Trump said the US opposes "any form of globalist regulation" of the most advanced AI.
In Westminster on Thursday, Plaid Cymru MP Ben Lake led a backbench debate urging binding rules for the most powerful systems. He said AI should be made safe by design, as aviation and pharmaceuticals are. Backbench debates do not create law, and no ministerial commitment from the debate had been reported by Friday.
The key signal is the government's formal reply to the committee. It will show whether ministers are willing to put frontier AI on a statutory footing, or will keep relying on voluntary testing deals.
Liability is a second open question. The committee says responsibility when safeguards fail is unclear, but that little can be settled until the Law Commission finishes its current review.
For investors, the takeaway is divergence. Washington is resisting international AI rules, while Britain's lawmakers are asking for the power to stop a model. Companies selling AI worldwide may soon face very different rulebooks, and different launch timelines, on either side of the Atlantic.
Written by AI from the public sources linked below.