OpenAI and Anthropic Back UK Lawmakers' Push for Binding AI Rules

OpenAI and Anthropic used a sweeping session of Britain's Parliament to make an unusual request of lawmakers: regulate us. Executives from both labs joined whistleblowers and national security figures at the hearing, arguing that voluntary commitments are no substitute for rules written into law. The message marks a shift from the industry's earlier preference for self-policing, and it lands as the UK weighs how far its AI Safety Institute's powers should reach.

The two companies have publicly supported mandatory safety testing for frontier models, including pre-deployment evaluations and reporting requirements for the most capable systems. Their appearance before MPs gives that position political weight, because the labs building the largest models are asking to be bound by rules rather than exempted from them. Whistleblowers at the same session described internal pressure to ship quickly, a claim that undercuts any argument that companies can police themselves.

Roughly a few hundred metres away, Britain's spy chief delivered a darker warning on the same day. He said AI could be used to develop new biological weapons, a scenario that moves the debate past chatbots and copyright and into national security. That framing matters for lawmakers deciding whether AI oversight belongs to a science ministry, a safety institute, or the security services.

The timing is not accidental. Britain has positioned itself as a middle path between the European Union's binding AI Act and the lighter-touch approach favored in Washington, hosting the first international AI safety summit in 2023. Its AI Safety Institute has focused on evaluation rather than enforcement, and legislation to give it statutory teeth has been repeatedly deferred. Pressure from the labs themselves removes one of the main objections to writing those rules.

For builders, the practical question is what mandatory testing would cover. Frontier-model evaluations, red-teaming for dangerous capabilities, and incident reporting are the likely core, with smaller open-weight developers possibly facing lighter obligations. Any UK regime that diverges from the EU's could raise compliance costs for teams shipping in both markets, while a closer alignment would make Britain a test case for the EU's rules.

What to watch next is whether the government converts this testimony into a bill with real penalties, and whether the security warning changes the scope from model safety to biosecurity controls on training data and access. The labs have said they want rules. The open question is whether they will accept the version Parliament writes.