OpenAI and Anthropic Push Britain to Make AI Safety Rules Binding

OpenAI and Anthropic told British lawmakers that voluntary AI safety pledges are no longer enough and that the U.K. should introduce binding rules for frontier models, sharpening a public split with Google DeepMind and Meta, which resisted calls for new legislation. The clash came during a parliamentary hearing that also heard warnings from intelligence officials about the growing danger of AI-enabled biothreats.

The exchange before the Business, Innovation, Science and Trade Committee in Westminster was more than a policy seminar. It became a rare public record of disagreement among some of the most influential AI companies in the world over whether self-regulation can still work as model capabilities accelerate.

For British lawmakers, the hearing underscored a central question now confronting governments: whether to continue relying on cooperation from powerful U.S.-based labs, or to require mandatory reporting, testing and oversight before advanced systems are deployed.

What British MPs heard from the frontier AI companies

British MPs heard sharply different views from executives representing two of the world’s leading frontier AI labs, and those differences went straight to the heart of how governments should govern advanced models.

Anthropic’s Europe chief, Liam Booth-Smith, was the most direct. Asked whether his company believed AI safety now required a legal framework, he said the rules should be binding and argued that voluntary promises are not sufficient on their own.

“The rules should be binding,” Anthropic’s Liam Booth-Smith told the committee, adding that the company has long believed voluntary commitments are not enough.

OpenAI vice president Tom Duff Gordon took a more measured tone, but still moved in the same direction. He backed a statutory framework and an international convention, while warning that some requirements should be mandatory at the outset, including incident reporting.

“There should be some elements that should be mandatory,” Duff Gordon said, adding that relying entirely on voluntary commitments for frontier-risk systems was not a good long-term strategy.

Taken together, the remarks suggested that at least some leading labs now see hard law as the next step in AI governance, especially as systems grow more capable and the potential consequences of mistakes widen.

Why this hearing mattered beyond Westminster

The committee hearing mattered because it brought the industry’s internal disagreement into the open in a second major jurisdiction, giving British policymakers a clearer view of where the fault lines lie.

Committee chair Liam Byrne said the aim was to illuminate the differences among companies that often appear aligned in public messaging about safety. In his view, the hearing showed that governments cannot assume firms will regulate themselves effectively without formal pressure.

That is especially significant in Britain, where ministers have sought to present the country as a bridge between the U.S., Europe and global AI governance debates. The parliamentary record now includes direct testimony from companies that favor binding oversight and others that prefer standards and existing institutions over fresh laws.

How Google DeepMind and Meta pushed back

Google DeepMind and Meta declined to endorse new legislation, signaling that the industry consensus on AI safety regulation remains far from settled.

Asked whether he supported a statutory framework, Google DeepMind’s Owen Larter steered MPs toward standards rather than new law.

“Focus on the standards is what we would be doing,” Larter said, stopping short of backing a legal framework.

Meta’s Rob Sherman argued that Britain already has a regulatory base to work from and suggested lawmakers should not behave as though they were starting from zero.

Sherman told MPs that Britain’s regulators are “not writing on a blank page,” pointing to the existing policy framework already in place.

The contrast was stark. On one side were companies saying voluntary promises are too weak for frontier systems; on the other were firms insisting that standards and existing regulators should do most of the work.

That divide is politically important. Britain has been positioning itself as a serious player in AI safety policy, but the hearing made clear that even the most advanced firms do not agree on what that should mean in practice.

Who wants binding AI laws and who does not?

OpenAI and Anthropic supported stronger legal obligations, while Google DeepMind and Meta did not explicitly back new legislation.

• OpenAI: supports a statutory framework and mandatory incident reporting.
• Anthropic: says binding rules are needed because voluntary commitments are insufficient.
• Google DeepMind: prefers focusing on standards rather than new laws.
• Meta: argues the U.K. already has regulators and rules to build on.

The hidden problem: Britain’s AI safety system is still voluntary

Britain’s current AI safety regime remains voluntary in practice, according to senior officials responsible for it.

That point emerged when lawmakers questioned civil servants from the AI Security Institute, the U.K. body that tests frontier models before they are released. The answer was blunt: the institute can only examine systems when companies decide to share them.

Cat Little, the Cabinet Office’s top civil servant, acknowledged the limits of the current setup when pressed by MPs. Henry de Zoete, who leads the AI Security Institute, described a system that depends on cooperation from companies and foreign governments and cannot compel access on its own.

De Zoete said the institute is reliant on companies and foreign governments for access to models, and he made clear that it does not certify systems or approve them for release.

He also rejected the idea that the institute acts like a formal safety stamp.

“What we’re not doing is kite marking,” de Zoete said, using the British term for a safety certification. “We’re not saying yes or no to any sort of model.”

That admission is central to the broader debate. Britain has built a respected technical institution, but the hearing showed that it lacks a legal mechanism to stop a company from releasing a model if officials believe it is too risky.

What the AI Security Institute can and cannot do

The institute can examine frontier models, but it cannot force companies to hand them over or block deployment on its own.

The structure reflects a broader tension in AI policy: technical expertise can improve oversight, but without legal powers, oversight can become dependent on the goodwill of the very companies being monitored.

How U.S. control complicates British oversight

Britain’s oversight model is complicated by the fact that the most advanced AI firms are American, and access to their systems often depends on decisions made in Washington.

OpenAI’s Duff Gordon said the company provides models to its U.S. counterpart before deployment and also shares them with the U.K. institute, but he tied that access to the policy environment in the United States.

He said OpenAI would continue providing models to the U.K. institute before release “as long as we’re able to do that under the U.S. government.”

That remark captured a defining reality of frontier AI oversight: Britain may want to shape safety checks, but the leverage over U.S. firms ultimately sits in the United States, where export controls, national security considerations and domestic regulation can all influence how much access foreign regulators receive.

Anthropic’s Booth-Smith also highlighted the issue by explaining that the company had at one point been unable to give the U.K. institute full unrestricted access to one of its models because of U.S. constraints, though he said access has since been restored.

The implication is clear. Even when Britain wants transparency, it can end up waiting for another country’s rules to allow it.

Why the whistleblowers’ testimony changed the tone

Former employees of leading AI labs told MPs that industry self-regulation is not trustworthy, adding a firsthand critique from people who had worked inside the companies under discussion.

Among them was Daniel Kokotajlo, a former OpenAI staffer, who said the companies cannot simply be trusted to police themselves.

Kokotajlo told MPs that he does not believe AI companies can be trusted to self-regulate.

Steven Adler, who left OpenAI in late 2024, said he was not confident in the industry’s safety culture. Geoffrey Irving, another former researcher, argued that the firms act as though they are trapped in a competitive race that governments cannot interrupt, but said that assumption is wrong.

Irving told the committee that the idea that companies cannot stop is false, arguing that they could simply choose to slow down or halt deployment if necessary.

The presence of whistleblowers mattered because it gave MPs a view beyond official company messaging. Their testimony suggested that internal pressures, competitive dynamics and incentives to move fast may be undermining safety commitments from within.

What former staff said about the race narrative

The former employees challenged a common industry defense: that each lab must keep shipping because competitors will otherwise surge ahead.

1. They said the race is not inevitable.
2. They argued companies can pause or slow deployments.
3. They suggested that external regulation may be needed to change incentives.

That message aligns with one of the committee’s central concerns: if every company believes it cannot afford to slow down, then voluntary restraint becomes structurally weak, even before any specific safety failure occurs.

MI5’s warning put the policy debate in a security context

The hearing took place at the same time Britain’s domestic intelligence chief was sounding the alarm about the misuse of AI for bioweapons, giving the policy discussion an unusually stark national-security backdrop.

MI5 director Ken McCallum warned that AI could significantly change the stakes around existing or future biological weapons threats. He said that, while such a scenario has not yet materialized, it is the risk he regards as most serious.

That warning sharpened the political tension surrounding the parliamentary hearing. If the country’s top domestic intelligence official is warning about potential misuse, MPs can reasonably ask whether a voluntary model review regime is robust enough.

Liam Byrne said the juxtaposition of MI5’s warning and the government’s voluntary AI approach showed, in his words, how much the U.K. still has to do to close the gap between risk and regulation.

For lawmakers, the issue is no longer abstract. It is about whether future models could help malicious actors design harmful biological agents, automate cyberattacks or produce other dangerous capabilities at scale.

Why Liam Byrne wants an international deal

The committee chair wants Britain to push for an international agreement on AI safety, and he sees the country’s diplomatic calendar as a possible lever.

Byrne said Parliament will spend substantial time examining how the U.K., as a middle power, can influence AI governance. He wants ministers to use Britain’s upcoming G20 presidency, alongside South Korea’s later turn in the chair, to broker a convention that does not depend on the current White House.

His logic is pragmatic. If frontier AI policy is being shaped by a small number of U.S. firms and by decisions in Washington, then Britain may need to build a broader multilateral coalition rather than rely on bilateral cooperation alone.

That would be an ambitious diplomatic project, but one that fits the scale of the problem. Advanced AI models are being developed internationally, deployed globally and used across sectors faster than any one national regulator can track.

What an international convention could include

A serious international agreement could, in theory, cover the kinds of obligations lawmakers discussed in the hearing.

• Mandatory incident reporting for serious model failures.
• Pre-deployment testing and evaluation requirements.
• Access for trusted national safety institutes.
• Rules on withholding or delaying release when models cross agreed risk thresholds.
• Information-sharing standards for cross-border AI safety coordination.

None of that is simple to negotiate. But the committee’s hearing showed growing pressure for governments to move beyond loose commitments and toward frameworks that have teeth.

The wider meaning for AI governance

The parliamentary session exposed a split that is likely to shape AI regulation well beyond the United Kingdom.

On one side are firms such as OpenAI and Anthropic, which increasingly acknowledge that frontier AI may require mandatory reporting and legal safeguards. On the other are companies such as Google DeepMind and Meta, which appear more comfortable with standards, existing regulators and voluntary collaboration.

That disagreement matters because it affects what governments think is feasible. If the companies themselves cannot agree, lawmakers may feel more justified in choosing a tougher route.

It also matters because the world is moving toward a period in which AI systems are expected to be more capable, more autonomous and more deeply embedded in critical infrastructure. In that environment, voluntary commitments can look fragile quickly, especially after the first serious incident.

Britain now faces a choice between being a testing ground for soft governance or a leader in binding oversight. The hearing suggested that the pressure for the latter is building, even as the state’s current powers remain limited.

Timeline of the hearing and key moments

The following timeline captures the most important developments that emerged during the committee session and the parallel security warning.

What happens next?

The immediate next step is political rather than technical. British lawmakers will continue examining whether the U.K. should keep relying on voluntary cooperation or move toward a binding regime for frontier models.

That discussion will likely focus on three issues: how to define high-risk systems, how to enforce disclosure and testing, and how much Britain can do independently when the leading AI labs are based overseas.

For now, the hearing has made one thing plain. The debate over AI safety is no longer only about best practices or public reassurance. It is about whether the world’s most powerful models can be governed by goodwill alone, or whether the next phase of AI will require law.

As Byrne put it, the risks are rising faster than the current safety architecture can keep up.