The UK work on pre-release testing for frontier AI models has stopped advancing, according to reporting this week, leaving labs, evaluators and the companies buying their systems without a published answer to a basic question: who checks a model before it ships, and under what power?

The programme lapsed after the Department for Science, Innovation and Technology was abolished in Andy Burnham's reorganisation of Whitehall. The department that commissioned the testing plans no longer exists under that name, and no successor programme has been described in the material published so far. Ministers still answer for AI policy in public. The machinery that was meant to give that policy teeth does not currently have a home.

What was supposed to be tested

Pre-release testing is the narrow, practical end of AI safety policy. Rather than regulating what a model says, it asks labs to hand over a frontier system to independent evaluators before deployment, so that dangerous capabilities can be probed while the developer can still change the weights or the guardrails. It is the part of the agenda that industry has been least able to dismiss as theoretical, because it costs money and creates a paper trail.

That plan was already lighter than the binding regime campaigners wanted. It is now the plan that nobody in government is publishing against, and the gap is visible to anyone trying to work out what a compliance calendar for 2027 looks like.

A minister without a department

The political context matters here. Burnham used his Liverpool conference appearance on Saturday to rule out an early general election and to talk up a 2029 horizon. Reorganisation is a governing choice, and the abolition of DSIT is now part of it. But the practical effect for AI firms is that the department that owned the testing question was retired before the question was answered.

Industry asks for rules it can plan around

Perhaps the least expected development is that parts of the industry are now the ones asking for legislation. The reporting indicates OpenAI has been pressing for a statutory framework, which is not the position a company takes when it expects to be left alone. Firms want certainty: a defined duty, a named regulator, a test they can pass and then point to.

The Ada Lovelace Institute published its own assessment of the UK's regulatory options on 25 September, setting out scenarios for how binding rules could be rebuilt and explicitly reopening the case for them. That is a think-tank contribution, not law, but it fills part of the vacuum the collapsed testing work left behind.

What regulation actually requires this month

On the evidence available, nothing new. There is no fresh statutory pre-release test obligation in force for frontier developers, no published evaluation regime with a named authority, and no bill whose passage would change either of those facts this month. What exists is a patchwork: sectoral rules where AI touches regulated activity, voluntary commitments from the largest labs, and enforcement that arrives after harm rather than before deployment.

So the Owen questions have unusually clean answers right now:

  • What was evaluated? Nothing new under a UK frontier testing scheme, because the scheme is not operating.
  • By whom? No successor body has been named.
  • What does UK regulation require this month? Existing duties and self-declared commitments, not a pre-release test.

That is a strange position for a country that spent two years telling the world it had found a middle path between Brussels and Washington: a light regime with a serious testing function attached. The testing function is the part that has come loose.

Where the numbers do move

Enforcement statistics elsewhere show what happens when the state does keep counting. City of London Police's Report Fraud for FY25-26 records AI-identified fraud reports up 395 per cent to 956, with associated losses rising from 1.2 million to 9.6 million pounds. Those are police-reported figures, and a rise in AI-identified cases also reflects better detection and classification rather than a one-for-one increase in underlying crime. Read with that caveat, the direction is still clear: fraud is where AI is already being policed by numbers.

Frontier models are not. Nothing in that fraud data tells you whether a model was safe before release, because nobody is publishing that test.

What to watch

Three things would change the picture: legislation that puts pre-release evaluation on a statutory footing, a named body with the staff and the legal power to run it, and a minister willing to own both the timetable and the failures. Until one of those appears, the UK's frontier AI testing regime is a plan without a department, and the industry asking for rules is asking a government that has not yet said who answers.