OpenAI, Anthropic, Google DeepMind and Meta tell British MPs they back outside safety testing, but split on how binding the rules should be

On 13 October, policy directors from OpenAI, Anthropic, Google DeepMind and Meta gave evidence to the House of Commons Business, Innovation, Science and Trade Committee. MLex reported that the firms broadly agreed on the need for outside scrutiny and safety testing but differed on how binding the rules should be, and that they broadly committed to give the UK AI Security Institute access to unreleased models where they are “legally allowed” to. The committee's own pages refused us, so this story rests on the outlets named.

OpenAI's Tom Duff Gordon told MPs the company had scrapped the release of GPT-6.1 Astra. “We found it didn't refuse enough, it hadn't progressed enough with alignment,” he said, according to the Press Association. He added: “It is evidence that OpenAI takes safety really seriously, that we are not racing to the bottom.”

Anthropic's Lord Booth-Smith said: “For our deployed models, we think they are safe and we have mitigations in place.” He also told the committee: “It is possible that real harm could be caused in the space of helping people to create biological or chemical weapons.” On 13 October the Switchboard also ran Britain's AI minister telling frontier labs not to release systems whose risks they cannot control.

• Reported The four labs broadly agreed on outside testing but differed on how binding the rules should be, MLex reported. MLex
• Claimed OpenAI's Tom Duff Gordon: “We found it didn't refuse enough, it hadn't progressed enough with alignment.” Press Association via The Irish News
• Claimed Anthropic's Lord Booth-Smith: “For our deployed models, we think they are safe and we have mitigations in place.” Press Association via The Irish News