OpenAI and Anthropic back binding AI safety rules at UK hearing

OpenAI and Anthropic backed binding AI safety requirements at a UK parliamentary hearing on October 13, while Google DeepMind and Meta declined to endorse new legislation. The divide had a concrete backdrop: British officials acknowledged that their oversight regime is voluntary, and its model-testing institute cannot stop a company from releasing a model it considers dangerous.

The Business, Innovation, Science, and Trade Committee heard from representatives of four companies developing the most advanced AI systems, four former employees turned whistleblowers, and senior civil servants. The session put both the industry’s preferred rules and the government’s current powers before a cross-party panel of MPs, according to WIRED’s account.

Binding duties versus standards

Anthropic’s head of Europe, Liam Booth-Smith, gave the clearest endorsement of legislation. Asked whether AI safety needed a statutory framework—rules established in law—he said companies should face binding obligations. He described that as Anthropic’s consistent position, rather than a new concession, and said voluntary commitments were insufficient.

OpenAI vice president Tom Duff Gordon supported a statutory framework and an international convention, but identified a narrower starting point. Some requirements should be mandatory, he said, beginning with incident reporting. He rejected reliance entirely on voluntary commitments as a sound long-term approach to risks from the most capable models.

Google DeepMind’s Owen Larter answered the same question by emphasizing standards rather than endorsing legislation. Meta’s Rob Sherman pointed to Britain’s existing regulators, saying MPs were “not writing on a blank page.” Neither backed a new statutory framework in those exchanges. Their responses emphasized standards and existing institutions, in contrast to Anthropic’s call for binding company duties.

We’re not saying yes or no to any sort of model.

Henry de Zoete, director of the UK AI Security Institute, speaking to the committee

Testing depends on access, not approval

The officials’ testimony drew a distinction between evaluating a model and authorizing its release. Britain’s AI Security Institute, or AISI, tests advanced models before deployment when developers provide them. Director Henry de Zoete said access depends entirely on cooperation from companies and foreign governments. Cabinet Office permanent secretary Cat Little confirmed that the current regime is voluntary.

Asked what would happen if a company released a model despite AISI judging it too dangerous, de Zoete acknowledged that the institute could not prevent that release. He said its work was not a safety certification. He also stressed Britain’s limited leverage over developers at the technological frontier, which he described as American or Chinese.

Company testimony showed how access can also depend on Washington. Duff Gordon said OpenAI supplies models before deployment to the US institute, CAISI, and would continue providing them to Britain’s AISI as long as the US government allowed it. Booth-Smith said Anthropic had previously been unable to give UK AISI full, unrestricted access to Mythos 5.1, referring that question to the US government. He added that AISI now had access.

Former researchers challenge self-regulation

Former lab researchers offered a sharper critique of leaving safety decisions with developers. Ex-OpenAI employee Daniel Kokotajlo said he did not trust companies to self-regulate. Steven Adler, who left OpenAI in late 2024, said he lacked confidence in the industry’s safety culture. Geoffrey Irving challenged the belief that labs are trapped in a race governments will not interrupt and that companies cannot leave. He said both beliefs were false.

Committee chair Liam Byrne told WIRED afterward that exposing differences between the companies was deliberate. The committee would devote substantial time to Britain’s options as a middle power in AI, he said. His proposed route was international: use Britain’s coming G20 presidency, alongside South Korea’s subsequent presidency, to pursue a convention that would not depend on the current White House.