20 November 2026
The Register's 'UK MPs Want AI Watchdog With Power to Pull the Plug on Dangerous Models' Has the Wrong Gap in It
The people who know exactly how dangerous these models are cannot tell you. That sentence does not appear in the Register's coverage of Guardrails for Growth: Making AI Safe1, published Tuesday. It should be the story.
By end of 2027, the Burnham government will pass something it calls binding AI safety rules. It won't give the AI Security Institute power to hold a release. I said that in October. The Commons Business, Innovation, Science and Trade Committee's report doesn't change it. Here's why.
1. The AI Security Institute cannot compel anyone to hand over a model.
Its work depends on voluntary cooperation.1 The committee says so. It wants a separate statutory regulator. Fine. Write down when that body opens its door and what happens to the models released before then.
2. Some released models cannot be recalled.
Liam Byrne told the press that some models, once out, cannot be pulled back by their developers.2 The power MPs want covers what hasn't happened yet. After deployment, that power addresses nothing. A halt authority over open-weight models already in the wild gives you a document, not a stop.
3. The people closest to the problem have NDAs.
This is the number the coverage buried. Former AI employees told MPs that restrictive contracts, commercial pressure and financial penalties discourage staff from raising safety concerns.1 MPs want mandatory incident reports and protected channels for whistleblowers.3
The proposed regulator gathers from outside. The employment contract stops flow from inside. The committee's answer to a contract problem is a new statutory body. Those solve different things, and the committee proposes only one.
4. Someone tried to build independent scrutiny already. The market said no.
Aengus Lynch, who worked with Anthropic on three research papers, told the committee he tried to launch a third-party AI evaluator because no body held labs to account. The idea didn't attract enough support.4 The committee now proposes a statutory version of exactly what Lynch couldn't fund. That may work. But MPs heard this in oral evidence and built recommendations around everything except the clause that stopped people speaking.
5. Anthropic, OpenAI, Google DeepMind and Meta all gave evidence.
The committee questioned representatives from all four alongside researchers and former employees.1 Former employees separately said they'd lost confidence in lab leadership.4 The lab reps sat in those sessions and watched. The gap between what each group could say tells you more about UK AI safety than any of the ten gaps the report names.
6. The recommendations bind nobody.
The government has not committed to the proposed legislation.5 The bill MPs want goes to the next parliamentary session.6 The G20 summit arrives in November 2027.6 The international compact the government wants to lead lands before the domestic statute it would need to back it up.
What everybody half-knows and nobody writes down: the safety knowledge MPs spent months trying to extract from outside these labs already exists inside them. An NDA stops it moving. The committee proposes one new body. It needs two things, and the other one amends employment law.