The report, Guardrails for growth: making AI safe, comes from the Business, Innovation, Science and Trade Committee, which spent months taking evidence from researchers, former lab employees, and representatives of OpenAI, Google DeepMind, Anthropic and Meta. All four companies assured MPs they would not release a model they considered unsafe. The problem, the committee says, is that they do not agree on what unsafe means — and British law does not generally require an independent opinion before shipping.

Committee chair Liam Byrne put it more sharply: “We cannot give today’s AI leaders unfettered freedom to behave like modern-day Dr Frankensteins, creating dangers beyond anyone’s control.” The report’s answer is to move safety from a promise into a duty — with a regulator that can act before a model ships, not just investigate afterwards.

What the committee wants#

The centrepiece is a new statutory regulator, separate from the AI Security Institute. The institute would keep testing frontier models while the new body enforces. Developers of designated high-risk systems would have to evidence their safety claims and give independent researchers access; significant changes to a model’s capabilities or safeguards would trigger fresh scrutiny.

The powers are what make this different from every previous British AI proposal. The regulator could intervene while models are still being developed: halt high-risk training and testing, block a deployment, or suspend a service if the risks cannot be contained. Alongside that, the committee wants mandatory reporting of incidents and near misses, and stronger legal protections for whistleblowers — after hearing that restrictive contracts, commercial pressure and financial disincentives discourage staff from speaking out.

A wooden judge’s gavel resting on a green parliamentary despatch box beside a glowing microchip with luminous circuit traces
Editorial illustration: the gavel meets the chip — the committee wants the law to reach inside the training run. AI-generated illustration.

Ten gaps, one theme#

The report identifies ten holes in Britain’s approach, and they share a theme: nobody has to tell anyone anything. AI firms face no general statutory duty to notify government of incidents they identify, so ministers may not even know something went wrong — let alone spot patterns in a model’s behaviour.

The report makes the point concrete in its footnotes, assembling a ledger of 2026 containment failures by name. Every entry was disclosed by the company responsible; none was filed with a British regulator, because there is no regulator to file one with and no obligation to file. The committee wants a framework with defined reportable incidents and near misses, standardised severity classifications, notification timelines, and clear rules on who reports to which authority.

The companies split on the fix. OpenAI and Anthropic supported placing incident reporting on a statutory footing. Google DeepMind and Meta backed standardised frameworks with clear routes and authorities instead — a form, in other words, rather than a duty with consequences attached. The committee sided with the statutory route.

What the witnesses said#

The most striking evidence came from people who had worked inside the labs. Daniel Kokotajlo, formerly a researcher in OpenAI’s governance division, told the committee he left because he lost confidence the company’s leadership would behave responsibly around the technology it was building. Steve Adler, another former OpenAI employee, described pressure not to use words such as “dangerous” when talking about AI, and to play up the benefits rather than the risks.

Geoffrey Irving, chief scientist at the government’s AI Safety Institute, told MPs he put the chance of AI destroying society at 50%. Aengus Lynch, who collaborated with Anthropic on alignment research, said he had tried to launch a third-party alignment evaluator because nothing held labs to account — and the idea failed to gain traction.

The report also looks beyond the frontier labs. It warns that Chinese developers are producing increasingly open-weight models that are cheap, easy to adopt and hard to control once released, and that systems built outside the UK could become widely embedded in the British economy. A regulator watching only the big domestic labs would be guarding the wrong door.

A large red emergency stop button on a wall in a dim server room, rows of blinking server racks fading into shadow
Editorial illustration: the power the committee wants — a regulator that can stop a system, not just audit it afterwards. AI-generated illustration.

The £400 billion argument#

The committee frames safeguards as a growth policy, not a brake on one. Using OECD productivity estimates, it calculates that AI could leave Britain’s economy up to £400 billion larger over the next decade — but only if the technology is trusted by individuals, businesses and public institutions. Stronger rules, the argument goes, are what get smaller firms to adopt AI in the first place.

MPs want the laws on the books in the next parliamentary session, and want Britain to use its 2027 G20 presidency to push other countries toward tougher rules on AI testing, incident reporting and risky releases.

What to watch#

None of this is law yet. The committee’s recommendations are not binding, and the government has yet to commit. Whether ministers will hand a watchdog the power to stop AI development — and define which systems count as high-risk — is the decision the report is really asking for.

The second is the government’s response to a similar call from a separate committee, the Joint Committee on Human Rights, which asked for a dedicated AI Act and a statutory body in September. If ministers answer that with a timetable, this report lands on a desk already moving; if not, the voluntary model survives another round.

Sources: The Register, “UK MPs want AI watchdog with power to pull the plug on dangerous models”; Computer Weekly, “UK government warned that lax AI safety undermines GDP boost”; Yeandel, “AI incident reporting: OpenAI, Anthropic, Google DeepMind, Meta evidence”.