OpenAI and Anthropic back mandatory AI rules in UK hearing

A split over binding rules

Anthropic’s head of Europe, Liam Booth-Smith, said voluntary commitments were not enough. OpenAI vice president Tom Duff Gordon was more cautious, but also supported mandatory rules and an international convention. He said some requirements, especially incident reporting, should be compulsory; relying on voluntary promises alone was not a sensible long-term plan for reducing risks from frontier models.

Google DeepMind’s Owen Larter said the company would focus on standards. Meta’s Rob Sherman pointed to existing UK regulators and said the country was not starting from scratch.

The committee’s chair, Liam Byrne, told WIRED that MPs wanted to show the breadth of views among some of the world’s richest companies and make clear that voluntary self-regulation could not be relied on.

The hearing took place as MI5 director Ken McCallum warned that AI could help create biological weapons. He said that scenario had not yet materialised, but the prospect of AI changing the scale of threats from existing or new biological weapons was his greatest concern.

Byrne connected McCallum’s warning to the parliamentary debate: while the MI5 chief was discussing AI’s potential to significantly increase bioterrorists’ capabilities, the Cabinet Office’s permanent secretary was effectively acknowledging that UK AI regulation remained voluntary.

Britain’s checks depend on access

The UK AI Safety Institute (AISI) tests frontier models before release, but only when companies choose to submit them. Its director, Henry de Zoete, said the institute depends entirely on cooperation from companies and foreign governments to access models. Cabinet Office senior official Cat Little confirmed that the system is currently voluntary.

Asked what happens if AISI considers a model too dangerous but its developer releases it anyway, de Zoete said the institute cannot stop the release. It does not issue a safety mark or decide whether a model can be launched. He added that frontier models are built by US or Chinese companies, limiting what Britain can do.

Access also depends on Washington. Duff Gordon said OpenAI gives models to the US Center for AI Standards and Innovation (CAISI) before release, and will keep sharing them with AISI before release as long as the US government allows it.

Booth-Smith said Anthropic could not at that point give AISI full, unrestricted access to Mythos 5.1, describing the limitation as a question for the US government. He said AISI now has that access.

What former employees want MPs to consider

Four former AI lab researchers also testified, including former OpenAI employee Daniel Kokotajlo. He told MPs he did not believe companies could be trusted to regulate themselves.

Stephen Adler, who left OpenAI at the end of 2024, said he was unsure about the industry’s safety culture. Jeffrey Irving said labs see themselves as part of a race that governments cannot stop and that they cannot leave. Both claims were wrong, he argued: companies could simply stop.

Byrne told WIRED the committee would spend considerable time discussing Britain’s future as a middle power in AI. He wants ministers to use the UK’s upcoming G20 presidency to pursue an international convention with South Korea, which will hold the presidency after Britain, independent of the current US administration. Byrne noted that the G20 presidency would last longer than Donald Trump’s US presidency, and said risks were growing faster than Britain’s safety system could contain them.

I think the hearing’s sharpest point was not that companies disagree about the rules, but that the existing system has no way to make access or compliance compulsory. The proposed international convention may address that gap, but the testimony leaves a more immediate tension: Britain is being asked to oversee models whose release it cannot prevent and whose testing access may depend on decisions made in Washington.