OpenAI and Anthropic tell UK lawmakers: regulate us

AI leaders joined security chiefs and whistleblowers at a wide-ranging parliamentary session, while Britain's spy chief warned a few hundred meters away that AI could be used to create new biological weapons.

The Business, Innovation, Science and Commerce Committee, a cross-party panel of MPs, held a meeting on Tuesday with four frontier AI companies, four former employees of the companies and senior UK civil servants.

Liam Booth-Smith, president of Anthropic Europe, told the committee when asked if the company was of the view that AI safety now needed a legal framework: "Laws should be binding." He said the company "consistently believes that there should be binding rules for companies and that voluntary commitments are not enough."

Tom Duff-Gordon, vice president of OpenAI, was more cautious, but still supportive, advocating a legal framework and an international convention. "There should be some element that should be mandatory," Duff-Gordon said, starting with incident reporting. "We don't think everything based on voluntary commitments for border risk is a good long-term plan," he said.

However, their competitors are not following. "Focusing on standards is something we will do," Google DeepMind's Owen Larter said when asked if he supported a legal framework. Meta's Rob Sherman pointed to existing UK regulators, telling MPs they were "not writing on a blank slate".

The meeting exposed the regulatory divide and put the conflicting approaches of AI leaders on the agenda of the UK Parliament. Exposing the split in the second jurisdiction was a key point for committee chairman Liam Byrne, who told WIRED after the hearing that the committee set out to "bring to light the wide differences of opinion between some of the wealthiest companies on earth and show that voluntary self-regulation cannot be relied upon."

As four executives gave evidence to the AI Safety Committee, just a few hundred meters away, the director of Britain's security service warned about the dangers of AI.

MI5 director Ken McCallum used his annual threat update to warn of the misuse of bioweapons as a significant threat from artificial intelligence. "We haven't figured it out yet, but the use case I want to place the most emphasis on is the potential for artificial intelligence to alter existing or even new bioweapon threats," McCallum says.

In response to the comment, Byrne told WIRED: "The fact that the head of MI5 is warning in one part of Whitehall about the development of incredibly dangerous capabilities for bioterrorists, and then in another part of Whitehall, the permanent secretary of the Cabinet Office is basically admitting that our regulatory regime for AI is whatever you voluntarily have to tell you."

In the final part of the evidence hearing, civil servants admitted to the committee how desperate the current system was. The UK's Artificial Intelligence Security Institute (AISI) tests frontier models before release, but only when companies decide to deliver them.

Henri de Zoet, director of the AISI, acknowledged that the institute relies heavily on the cooperation of foreign companies and governments to gain access. Cat Little, a senior civil servant at the Cabinet Office, which houses the AISI, was pressed on whether the entire regime was voluntary. "Right now, yes, it is," he said.

When asked what would happen if AISI deemed a model too dangerous, but a company released it anyway, Dezoet acknowledged that AISI had no power to prevent it from doing so.

"What we don't do is mark the kite," de Zoet said, using the English term for a safety certificate. "We don't say yes or no to any model." He added: "It's either American or Chinese companies that are building across the border... There's only so much we as Britain can do about this."

Border managers also clarified that UK access would pass through Washington. "We are providing a pre-deployment of our models for CAISI," OpenAI's Duff Gordon said, referring to AISI's American counterpart. "We will continue to provide our pre-deployment models to AISI in the UK ... and will continue to do so as long as we can do so under the US government."

The tension of existing arrangements was made clear to the British authorities by Anthropic's Booth-Smith. "Unfortunately, we were not able to provide full and unrestricted access to [Mythos 5.1] to British AISI at the time... That's a question for the US government," he said, before adding that AISI now has access.

The session also heard from four former researchers at the lab, including former OpenAI employee Daniel Coccotaglou. "I don't think his companies told the MPs that self-regulation can be trusted. Steven Adler, who left OpenAI at the end of 2024, said he "doesn't trust the safety culture across the industry." Geoffrey Irving said labs believe they are locked in a race that governments won't interrupt and can't leave. "Both of these are lies," he said. "They can only stop."

After the hearing, Byrne told WIRED that his committee will now spend a lot of time examining the way forward for the UK as a middle power in AI. He wants ministers to use Britain's upcoming chairmanship of the Group of 20, alongside South Korea, which he is seeking to chair, to broker an international convention independent of the current White House. He said: "The presidency of the Group of 20 will stay in America longer than during Trump's presidency.

"The risks are increasing faster than our safety regime can handle," he added.