TL;DR
Senior ministers in Keir Starmer’s government had begun work on an AI safety law, including whether frontier labs could be made to submit models for testing before launch. The Guardian reports that work was lost in the turmoil of his last months. Andy Burnham then abolished DSIT, the department doing the work, and MPs are now pressing for the question to be reopened.
What was on the table
According to people briefed on the plans, officials reviewed which existing powers ministers already held and what new legislation could add. The central idea was compulsory pre-release evaluation for the most capable models. In June the government had also committed £115m to an incident response centre for AI systems that go rogue, and to a biosecurity programme.
Burnham’s early decision to scrap DSIT moved AI minister Kanishka Narayan, now promoted, into the Cabinet Office without a department behind him. Matt Clifford, who advised both Sunak and Starmer, warned on X that the reorganisation “wastes time and energy”. Dame Wendy Hall said “the fragmentation we have now has had a much worse impact” than DSIT’s bloat.
Why it has come back
The pressure has built quickly. This site reported Anthropic releasing its latest model without AISI evaluation and the King warning AI chiefs about existential dangers. Replying to Liam Byrne’s committee, AISI head Henry de Zoete said no one outside the US had seen Anthropic’s model before release, and that his teams could still study it afterwards.
Public opinion has moved too. In YouGov’s latest polling, slightly more than a third of Britons now place AI third among the things that could wipe out humanity, after climate breakdown and nuclear war. That share has more than doubled in a year. Narayan wrote in the Times that “nothing is off the table”.
The options being argued over
Chi Onwurah, chair of the Commons science and technology committee, suggested barring sales of commercial models whose frontier systems have skipped the safety institute, and making developers legally liable when an agent attacks a third party. Byrne argues that “a safety regime which is set and enforced by the UK, Europe, India, Canada and Japan would be robust”. Darren Jones wants Burnham to raise it at next year’s G20, which Britain will host.
Looking forward
Unusually, part of the industry is asking for rules: OpenAI has already urged Westminster to legislate, and this week its European policy head, Tom Duff Gordon, said the company backs tougher UK rules for the few firms building the most powerful systems. The difficulty is now institutional rather than political. A pre-release testing law needs a department to draft it, and the one that began the work no longer exists. UK businesses buying frontier models should assume the voluntary arrangement remains in place for some time yet.