OpenAI and Anthropic urge UK lawmakers to set binding AI safety rules

Representatives from OpenAI and Anthropic told a British parliamentary committee on Tuesday that voluntary safety pledges are no longer sufficient and that lawmakers should pass binding rules for frontier AI companies. Their testimony split the industry on the public record, as Google DeepMind and Meta declined to back new legislation at the same hearing.

The Business, Innovation, Science, and Trade Committee, a cross-party group of MPs, heard from four frontier AI firms alongside four former employees turned whistleblowers and senior civil servants. Anthropic’s head of Europe, Liam Booth-Smith, told the committee his company has “consistently believed that there should be binding rules on companies and that voluntary commitments are not enough.” OpenAI vice president Tom Duff Gordon was more guarded but still backed a statutory framework and an international convention, saying mandatory incident reporting should come first. “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 follow. Asked directly whether he supported binding rules, Google DeepMind’s Owen Larter said the company’s focus was on technical standards instead. Meta’s Rob Sherman pointed to Britain’s existing regulators, telling MPs the country was “not writing on a blank page.”

Committee chair Liam Byrne said exposing that divide was the point of the session, meant to show “that you couldn’t rely on voluntary self-regulation.” The hearing ran alongside a separate warning from MI5 director Ken McCallum, who used his annual threat update to flag AI’s potential to assist bioweapon development as a serious risk. Byrne said the contrast showed British policy lagging the threat: one part of government was warning about dangerous capability gains for bioterrorists while another admitted the regulatory regime stayed voluntary.

A voluntary regime

Civil servants confirmed how limited current oversight is. The UK’s AI Security Institute, AISI, tests frontier models for safety risks before release, but only when companies agree to hand them over. Its director, Henry de Zoete, said the institute cannot stop a company from releasing a model it judges too dangerous and does not issue any form of safety certification. Cabinet Office permanent secretary Cat Little confirmed the regime is voluntary “at the moment.” OpenAI’s Duff Gordon said the company shares models with Britain’s institute only as long as it is permitted to do so by the US government, which runs the equivalent American body, CAISI.

Four former lab staff also testified. Ex-OpenAI researcher Daniel Kokotajlo said he does not believe companies can be trusted to self-regulate. Former OpenAI employee Steven Adler said he was not confident in the industry’s safety culture. Geoffrey Irving, also a former lab researcher, rejected the idea that companies are locked into a race they cannot leave, saying: “They could just stop.”

Byrne told Wired his committee will next examine how Britain, as a middle power, might use its upcoming G20 presidency and South Korea’s term after it to push for an international AI convention that does not depend on the current US administration.