OpenAI and Anthropic back binding AI rules at UK Parliament hearing

OpenAI and Anthropic backed statutory AI-safety requirements during a UK parliamentary hearing on Tuesday. Google DeepMind and Meta stopped short of supporting new legislation. The split put the positions of four frontier-model companies on the record as MPs examined whether voluntary testing and reporting provide enough oversight.

At issue is whether some safeguards should become legal duties. Before the session, the Business, Innovation, Science and Trade Committee asked whether pre-release safety testing should be mandatory and whether companies should have to report serious AI incidents, including deceptive behavior, bypassed safeguards, unauthorized replication or evidence that human control may be failing. It also asked which international requirements should be compulsory.

Contemporaneous reports of the testimony described a clear divide. Anthropic policy chief Liam Booth-Smith supported a statutory safety framework and said rules for companies should be binding. OpenAI’s Tom Duff Gordon also backed a statutory framework and an international convention. He identified incident reporting as an initial element that should be mandatory rather than leaving frontier-risk governance entirely to voluntary commitments.

The same reports said Google DeepMind senior director Owen Larter did not endorse a statutory framework, emphasizing standards instead. Meta global policy vice president Rob Sherman pointed MPs to existing UK regulators and said policymakers were not starting from a blank page. The hearing established the companies’ policy positions but produced no new law or implementation timetable.

The debate focused on the limits of the UK’s current oversight system. Parliament’s hearing page says the AI Security Institute can evaluate risk when a developer grants it access to a model, but does not necessarily have the power to compel disclosure or stop a release. AISI says much frontier-developer risk management — including sharing models with third parties for pre-release evaluation — remains voluntary, and says it is not seeking to mandate requirements.

AISI’s published evaluation policy also says the institute is not a regulator, does not designate systems as safe and is not responsible for developers’ release decisions. Its evaluations act as a secondary check. Even so, its current technical agenda lists detecting, responding to and reporting incidents involving autonomous AI agents as a priority, including principles for reports to affected parties, governments and the public.

Separate contemporaneous reporting said OpenAI and Anthropic representatives would act on demonstrable evidence from AISI or UK security agencies that a model was unsafe. Both qualified that commitment: they could halt a deployment only while they still controlled the model, which may not be the case when another party hosts it.