OpenAI and Anthropic Tell British Lawmakers: Regulate Us

The Business, Innovation, Science, and Trade Committee held an evidence session with four frontier AI firms, four former staff members turned whistleblowers, and senior British civil servants. Anthropic's head of Europe, Liam Booth-Smith, advocated for binding rules for AI safety, stating that voluntary commitments are insufficient.

OpenAI vice president Tom Duff Gordon expressed cautious support for a statutory framework and an international convention, emphasizing the need for mandatory incident reporting. Google DeepMind's Owen Larter dismissed the idea of a statutory framework, focusing instead on setting standards. Meta's Rob Sherman noted that Britain's existing regulators were "not writing on a blank page."

The committee aimed to highlight the significant differences in opinions between the wealthiest AI companies, as chair Liam Byrne explained to WIRED. While the executives provided evidence on AI safety, British security service director Ken McCallum disclosed the danger of AI facilitating bioweapon misuse. McCallum's comment prompted select committee chair Liam Byrne to remark on the stark contrast between MI5's warning and the voluntary nature of the UK's regulatory regime.

Civil servants admitted that the current system relies solely on companies' cooperation for testing and assessing AI models before release. Henry de Zoete, director of the AI Security Institute (AISI), acknowledged the institute's reliance on voluntary cooperation from companies and foreign governments. The final part of the session saw civil servants confirm that the current regime is, indeed, voluntary.

AISI's Henry de Zoete clarified that the institute cannot prevent companies from releasing dangerous models, as these companies are primarily American or Chinese.

The liaison between British officials and the frontier AI firms was made clear by OpenAI's Duff Gordon, who stated that they provide predeployment models to the US counterpart of AISI. Anthropic's Booth-Smith revealed that they were unable to provide full unrestricted access to a particular model to UK AISI at the time, but AISI now has access. Former OpenAI staff member Daniel Kokotajlo and ex-OpenAI employee Steven Adler expressed their doubts about the companies' trustworthiness in self-regulating AI safety.

The session concluded with the select committee chair, Liam Byrne, outlining the upcoming focus on the UK's path forward in AI regulation, backed by Britain's upcoming G20 presidency alongside South Korea. Byrne emphasized the necessity of an international convention that does not depend on the current US leadership.