No Sanctions, No Teeth: UK's Voluntary AI Safety Regime Fails Its First Real Test

On Tuesday afternoon, First Secretary of State Louise Haigh stood before the TUC Congress in Brighton and delivered the British government’s most sober public assessment of artificial intelligence to date. AI, she said, has “enormous potential” to transform public services, strengthen businesses and deliver scientific breakthroughs — but there are “clearly huge risks to our national security and society if the right guardrails are not put in place.”

“We have a solemn duty to mitigate the risks and ensure that it is for the benefit — not the detriment — of our society,” Haigh told delegates, adding that ministers must “heed the warnings of those who are at the forefront of developing this technology” and be “ready to work with international partners” to prioritise public safety.

Hours later, The Guardian revealed just how large the gap between that rhetoric and Britain’s actual enforcement machinery has become: Anthropic will face no sanctions after declining to submit its latest model for pre-release testing by the AI Security Institute (AISI) — the first time a major frontier lab has simply opted out of the UK’s voluntary evaluation regime, and the clearest demonstration yet that “voluntary” is doing all the work in the phrase “voluntary safety arrangements.”

The scoop that framed the week

The story began on 8 September, when the Financial Times reported that Anthropic had withheld its newest model from AISI’s pre-deployment evaluation pipeline — the first such refusal by a major developer since the institute was founded in 2023 (and renamed from “Safety” to “Security” Institute in early 2025). Participation in AISI testing is not mandated by any UK statute. It rests on memoranda of understanding and the goodwill of the labs.

When goodwill ran out, the institute had nothing to impose. No fine. No suspension. Not even a formal rebuke. The Guardian’s revelation that no sanction will follow turns an obscure procedural footnote into a constitutional question about the UK’s entire approach to frontier AI: what exactly is a safety regime that cannot compel a test, and cannot punish its absence?

The timing is what makes the episode so uncomfortable for the company at its centre. Over the past fortnight, Anthropic has been the industry’s loudest advocate for restraint. CEO Dario Amodei’s essay “We Must Pace the Frontier” (published 12 September) called for a coordinated slowdown in capability gains across the frontier labs — a position publicly backed by OpenAI, Google DeepMind and Elon Musk. Co-founder Jack Clark went further, telling the BBC that a third-party AI “kill switch” may eventually need to be mandatory for companies, something society “might want to eventually pass rules around.”

A lab campaigning for binding rules on everyone else, while declining a voluntary check itself, is the sharpest illustration yet of why the voluntary version of frontier governance struggles to survive contact with competitive pressure.

A government talking in both directions

Haigh’s speech was only one voice in a cabinet that spent the week pointing in different directions. Business Secretary Jonathan Reynolds, speaking on BBC Radio 4’s Today programme, urged against both complacency and being “hyperbolic.” Asked about Clark’s kill-switch proposal, he replied: “I don’t think that’s a particularly helpful way to think about how we manage the risks — I’m not really sure what that would mean in practice.” Reynolds also warned that overregulation could cost the UK access to models developed abroad, which he argued would leave the country “less safe,” not more.

That echoes the Cabinet Office’s formal rejection — reported last week — of the cross-party kill-switch amendment, on the grounds that Britain “cannot simply turn AI off” and that blocking models domestically would not prevent their development or misuse elsewhere.

The tension is structural. The government wants to be the world’s most credible voice on AI safety and the most attractive European destination for AI investment. Those goals have coexisted comfortably while compliance was voluntary and free. The Anthropic refusal forces the choice into the open.

Parliament starts pulling levers

Westminster’s response has been swift, if procedural. Liam Byrne, the Labour MP who chairs the Commons business and trade committee, has written to Henry de Zoete, the head of AISI, requesting that he testify to Parliament. Byrne’s letter cites growing public concern about “unfettered, inadequately governed artificial intelligence development” and questions “the adequacy of current AI safety governance.”

His intervention follows the Joint Committee on Human Rights, which last week called for a statutory AI framework in the UK after a report found the technology responsible for numerous abuses of human rights. “Nowhere in the world, including the UK, has a current legislative and regulatory approach to AI that is fit for purpose,” said Labour MP and committee member Alex Sobel.

Even OpenAI — which has its own complicated history with pre-deployment commitments — has reportedly urged ministers to legislate now, while the political window is open, with binding rules aimed narrowly at the handful of frontier labs rather than the broader startup ecosystem.

The most consequential proposal comes from Darren Jones, formerly chief secretary to the Treasury, who wants Prime Minister Andy Burnham to place international AI regulation on the agenda when the UK hosts the G20 summit next year. Jones, though, attached a warning that will resonate with anyone who has watched Britain’s regulatory history: “The thing I would be worried about is if we go into a classic Westminster approach of creating a new regulator and inadvertently kill off the AI ecosystem in this country.”

Unions put workers into the frame

If the ministers spoke about national security, the TUC Congress in Brighton spoke about jobs — and provided some of the week’s most concrete testimony. Delegates passed a motion urging the government to introduce a national AI plan and to resist the influence of “the big tech lobby.”

TUC general secretary Paul Nowak called AI “a clear and present danger to workers and wider society” and urged Burnham to use the G20 presidency to drive global regulation that ensures the technology “benefits everyone, not just the super-rich.” Unite’s general secretary Sharon Graham said workers and public safety were “being completely lost in the AI debate.”

The specifics were less abstract. Teachers described AI systems already being used in classrooms to record and assess teacher performance. A Unite representative described crane operators being recorded and analysed to train models that will automate their jobs, and HGV drivers living under constant surveillance. “We have finance and IT workers whose employers see them as a cost to be saved, and AI as the tool to do it,” said Unite’s Pat Dowling.

Answering questions after her speech, Haigh signalled which side of that debate the government intends to occupy: “Where AI makes work more productive, then benefits should be felt by workers, not just shareholders or tech bros.”

The wider storm

None of this is happening in a vacuum. The UK debate is unfolding against the most extraordinary fortnight of insider warnings the AI industry has ever produced: Anthropic researcher Evan Hubinger’s assessment that there is a greater-than-10% chance AI “could kill all humans” within a decade; Jacob Coxon’s public resignation from Anthropic warning about the “fate of humanity in the next two years”; Bilal Chughtai’s departure from Google DeepMind and his warning late Monday that “AI has the potential to kill us all”; and OpenAI researcher Dan Selsam’s claim that capable models will “likely convince people that everything is fine” and “argue convincingly that humans should trust them with power.”

Meanwhile, US President Donald Trump spent Tuesday dismissing the entire genre, calling fears that AI will destroy humanity “a HOAX” and insisting the only guardrails needed are “a strong and smart” president — a reminder that whatever framework Britain builds, it will be building it largely alone among Anglosphere allies.

What happens next

The immediate watch-points are procedural but consequential. Whether AISI’s chief executive appears before Byrne’s committee; whether the government converts any of this week’s rhetoric into an AI Bill with actual enforcement powers; and whether Burnham embraces Jones’s G20 gambit or leaves it to quieter diplomatic channels.

The deeper lesson was already delivered, by accident, through the Anthropic episode. Britain’s frontier safety regime has now been tested exactly once under pressure, and the result is a matter of public record: a lab said no, and nothing happened. As one Westminster observer put it in the wake of the Guardian’s report, the affair is “a plain demonstration of what voluntary means” in Britain’s safety arrangements.

Haigh is right that the warnings should be heeded. The question her government now faces is whether heedance, on its own, is a policy.