Whether AI safety should rely on companies’ promises or binding rules became a point of open disagreement at a British parliamentary hearing. OpenAI and Anthropic backed some form of statutory framework, while Google DeepMind and Meta pointed instead to standards and existing regulators.
The evidence session also highlighted a practical limit in Britain’s current approach: its AI Security Institute can test frontier models before release only when it can obtain access. That access depends on cooperation from companies and foreign governments.
Companies diverge on binding rules
The Business, Innovation, Science, and Trade Committee heard from four frontier AI firms, four former employees who had become whistleblowers, and senior British civil servants. The cross-party panel of MPs asked whether voluntary commitments could provide enough protection as AI capabilities advance.
Anthropic’s head of Europe, Liam Booth-Smith, said the company supported binding rules. “The rules should be binding,” he told the committee, adding that voluntary commitments were not enough.
OpenAI vice president Tom Duff Gordon was more measured, but also supported mandatory measures. He backed a statutory framework and an international convention, saying that incident reporting could be among the elements required by law. “We don’t think basing everything on voluntary commitments for frontier risk is a good long-term plan,” he said.
Google DeepMind and Meta did not endorse new legislation. Asked about a statutory framework, Google DeepMind’s Owen Larter said the focus should be on standards. Meta’s Rob Sherman pointed to Britain’s existing regulators, telling MPs they were “not writing on a blank page.”
The disagreement put different views from major AI firms on the parliamentary record. Committee chair Liam Byrne told WIRED that revealing those differences was deliberate: he wanted to show how far apart some of the wealthiest companies were and why voluntary self-regulation could not simply be taken for granted.
Britain’s testing system relies on cooperation
The hearing turned from what rules companies support to what the UK can currently require. Britain’s AI Security Institute, or AISI, tests frontier models before release, but companies decide whether to provide them for testing.
Henry de Zoete, AISI’s director, said the institute relied entirely on cooperation from companies and foreign governments to gain access. Cat Little, the top civil servant at the Cabinet Office, which houses AISI, was asked whether the regime was voluntary. “At the moment, yes, it is,” she replied.
The committee also explored what would happen if AISI judged a model too dangerous but a company released it anyway. De Zoete acknowledged that the institute could not stop the release. Its role is not to certify models as safe or unsafe: “What we’re not doing is kite marking,” he said.
That leaves a gap between testing and control. A safety assessment can inform understanding of a model, but the institute has no power to prevent a company from releasing it. De Zoete also pointed to the international nature of frontier AI development, saying that many companies building at the frontier are American or Chinese and that there is only so much the UK can do.
Access can depend on decisions in Washington
Evidence from the company representatives showed that British access to models can also hinge on US government decisions. Duff Gordon said OpenAI provides predeployment models to CAISI, AISI’s US counterpart, and would continue providing them to the UK institute as long as it could do so under the US government.
Booth-Smith described a case in which Anthropic had not been able to give UK AISI full unrestricted access to [Mythos 5.1]. He said that was a question for the US government, and added that AISI now had access.
The exchange illustrated how national testing capacity and international dependencies intersect. Britain may have an institute to assess models, but access to a particular system can depend on the decisions of the companies developing it and the governments that oversee them.
Safety concerns and a push for international action
While the committee heard evidence on AI safety, MI5 director Ken McCallum warned in his annual threat update about potential bioweapon misuse. He described AI’s potential to change the stakes around existing or new bioweapons as the misuse case to which he would attach the greatest weight, while saying it had not yet been realized.
Byrne drew a connection between that warning and the voluntary nature of the UK’s current AI regime. He told WIRED that the contrast between concern about dangerous capabilities and a system that companies are not compelled to participate in showed why the committee was examining regulation.
Former researchers also challenged the idea that companies could safely regulate themselves. Daniel Kokotajlo said he did not think companies could be trusted to self-regulate. Steven Adler said he was not confident in the industry’s safety culture, while Geoffrey Irving argued that companies were not helplessly trapped in a race: “They could just stop.”
Byrne said the committee would spend significant time considering the UK’s path as a middle power in AI. He wants ministers to use Britain’s coming G20 presidency, alongside South Korea’s following presidency, to seek an international convention that does not depend on the current White House. His warning was blunt: “The risks,” he said, “are multiplying faster than our safety regime can control.”