Anthropic did not submit its most powerful artificial intelligence model for safety testing conducted by the UK government, according to a report from The Times. The disclosure puts a spotlight on the voluntary nature of AI safety evaluations in Britain and whether leading labs are fully cooperating with oversight efforts designed to manage risks from frontier AI systems.
What the UK Testing Program Covers
The UK's AI Safety Institute, established in late 2023, was created specifically to evaluate advanced AI models before and after public release. The institute relies on voluntary cooperation from AI developers, meaning it has no legal authority to compel companies to submit models for review. Anthropic had previously signaled support for government safety initiatives, making the omission notable. The company has publicly positioned itself as a safety-focused lab, and its decision to hold back its flagship model sits uneasily with that reputation.
Key Facts
- The UK AI Safety Institute depends on voluntary model submissions from AI developers
- Anthropic did not submit its most capable model for evaluation, per The Times
- The institute was set up following the 2023 UK AI Safety Summit at Bletchley Park
- Major AI labs, including OpenAI and Google DeepMind, have previously cooperated with UK testing
- Anthropic has not publicly explained which model was withheld or why
The report does not specify precisely which model was kept back, though it is understood to refer to the most capable system in Claude's model family at the time of testing. That gap matters because the whole point of the institute's work is to assess the models most likely to pose novel or serious risks. Evaluating an older or less capable version gives regulators an incomplete picture.
The voluntary framework only works if companies participate in good faith. Submitting a less capable model, or no model at all, makes the entire exercise less meaningful.AI policy researchers, as characterised by The Times
Voluntary Commitments Under Scrutiny
This development comes amid broader questions about how much weight voluntary safety commitments from AI companies actually carry. In 2023, several leading labs signed pledges at the Bletchley Park AI Safety Summit, agreeing in principle to cooperate with government evaluations. Those pledges lacked enforcement mechanisms, and the Anthropic situation suggests the gap between stated commitments and actual practice can be significant. Questions about the company's approach to safety have surfaced before, and coverage of Anthropic's outsized influence in the AI industry has only intensified scrutiny of its decisions.
The UK is not alone in grappling with this problem. The United States, the European Union, and other governments are all working to build evaluation infrastructure for frontier AI, and most face the same core tension: they need cooperation from the very companies they are trying to oversee. Legislation that would give regulators binding authority over model submissions has moved slowly in most jurisdictions.
For now, the UK AI Safety Institute continues to operate on goodwill. Whether this episode prompts a renegotiation of how labs engage with the institute, or spurs calls for stronger statutory powers, remains to be seen. Anthropic has not issued a public statement addressing the specific claims in The Times report. The company's silence leaves open the question of whether this was a deliberate policy choice or a matter of timing and logistics around a model release cycle.
What is clear is that the incident has handed critics of voluntary AI governance a concrete example to point to. If labs pick and choose which models face external scrutiny, the credibility of the entire safety testing ecosystem is weakened. Policymakers in Westminster are likely to revisit the terms under which AI developers engage with the institute, and this story may accelerate that conversation.
