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

OpenAI and Anthropic representatives told a British parliamentary committee that voluntary safety commitments are no longer sufficient and that binding regulations are needed to manage frontier AI risks. Anthropic’s head of Europe, Liam Booth-Smith, stated that the 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 echoed this sentiment, supporting a statutory framework with mandatory elements such as incident reporting.

Google DeepMind and Meta declined to endorse new legislation. Google DeepMind’s Owen Larter emphasized a focus on standards, while Meta’s Rob Sherman pointed to existing UK regulators, arguing that the country is ‘not writing on a blank page.’ The divergence was deliberate, according to committee chair Liam Byrne, who said the hearing aimed to highlight ‘wide differences in opinions’ among the world’s wealthiest AI firms and to challenge reliance on voluntary self-regulation.

During the same parliamentary session, Britain’s security service director warned that AI could enable the creation of new bioweapons. MI5 director Ken McCallum, speaking at a separate event nearby, described bioweapon misuse as a significant AI-related threat, stating that while not yet realized, the potential for AI to ‘change the stakes on existing or even new bioweapons’ is a key concern.

The hearing also revealed limitations in the UK’s current AI oversight regime. Britain’s AI Security Institute (AISI) tests frontier models only when companies voluntarily provide access, and lacks authority to block releases or certify safety. Henry de Zoete, AISI director, acknowledged the institute’s reliance on corporate and foreign government cooperation, noting that AISI does not ‘kite mark’ or approve models. Cat Little, a senior civil servant at the Cabinet Office, confirmed the regime’s voluntary nature, and de Zoete admitted AISI could not prevent a company from releasing a model even if it judged it too dangerous.

The session included testimony from four former AI lab researchers, including ex-OpenAI staffer Daniel Kokotajlo, who argued that companies cannot be trusted to self-regulate. Steven Adler, a former OpenAI researcher, expressed ‘not [being] confident in the safety culture across the industry,’ while Geoffrey Irving criticized the industry’s narrative of an unbreakable race that governments cannot disrupt.

Committee chair Byrne indicated the panel would prioritize examining the UK’s regulatory path forward, including using Britain’s upcoming G20 presidency to broker an international convention independent of the US. Byrne emphasized that risks are ‘multiplying faster than our safety regime can control,’ and suggested the G20 process could outlast the current US presidential term.