Anthropic and OpenAI Back AI Laws as Rivals Stop Short

Anthropic and OpenAI told British MPs that frontier AI safety needs mandatory rules. Google DeepMind and Meta declined to endorse a new statutory framework, exposing a policy divide among four companies whose cooperation underpins government testing.

According to WIRED’s account of the hearing, Anthropic explicitly backed binding safety rules, while OpenAI supported a statutory framework with mandatory elements beginning with incident reporting. Google DeepMind emphasized technical standards; Meta pointed lawmakers toward Britain’s existing regulators.

The practical stakes emerged in testimony from government officials. The UK’s AI Security Institute, known as AISI, depends on voluntary cooperation to obtain models for evaluation. Its director acknowledged that the institute cannot prevent a company from releasing a model it considers dangerous.

Those are related but separate questions: whether companies must report incidents, whether government testers can demand access, and whether an adverse evaluation can affect a release. The hearing showed support for regulation without establishing agreement on how far any of those powers should extend.

Anthropic Backed Binding Rules; OpenAI Named a Starting Point

Anthropic’s position was the most explicit. Asked whether AI safety needed a statutory framework, its head of Europe, Liam Booth-Smith, told the committee:

“The rules should be binding.”

Booth-Smith said Anthropic had consistently supported binding rules for companies and considered voluntary commitments insufficient. That is a clear endorsement of legal obligations, rather than a request to improve the industry’s existing pledges.

OpenAI’s vice president, Tom Duff Gordon, also supported a statutory framework and an international convention. His account was more specific about an initial requirement: incident reporting.

“We don’t think basing everything on voluntary commitments for frontier risk is a good long-term plan,” Duff Gordon said.

The distinction between the two positions matters. Anthropic stated a broad principle. OpenAI identified an activity that should become compulsory. Neither statement, as reported, establishes agreement on a complete regulatory system, including which models it would cover, what enforcement powers it would create, or what penalties would apply.

Mandatory incident reporting would address one information gap: whether authorities learn about qualifying failures or dangerous events. Its effectiveness would depend on definitions that the reported testimony does not settle, such as what counts as a reportable incident and how quickly a company must disclose it.

It would not, by itself, guarantee government access to an unreleased model or give officials authority to delay deployment. Treating those powers as interchangeable would overstate what OpenAI endorsed.

The meaningful development is narrower but still important: two major frontier labs publicly accepted that at least some safety obligations should no longer depend on company discretion.

Google DeepMind and Meta Offered Different Alternatives

Google DeepMind’s Owen Larter did not follow Anthropic and OpenAI in endorsing a statutory framework. Asked directly about one, he said:

“Focus on the standards is what we would be doing.”

Meta’s Rob Sherman instead emphasized the regulatory institutions Britain already has, telling MPs they were “not writing on a blank page.”

These responses should not be flattened into a claim that either company opposes all AI regulation. Google DeepMind emphasized standards as the route forward. Meta’s response directed attention to existing regulators rather than a new statutory framework. Those are distinct positions, even though both stopped short of the endorsement the committee sought.

Technical standards and binding law are not inherently competing choices. A standard can describe how to conduct an evaluation or document a risk. Legislation can determine whether an organization must follow that standard and what happens if it does not.

The unresolved question is therefore not simply whether standards are useful. It is whether participation should remain discretionary, and whether existing authorities can secure the information and cooperation they need.

WIRED’s account does not establish a separate answer from every company on mandatory incident reporting or compulsory testing access. The clearest four-way comparison concerns their willingness to back a new statutory framework. It would go beyond the evidence to assign each lab a detailed position on every potential obligation.

AISI Can Test Models, but It Cannot Block Their Release

The officials’ testimony exposed the limits of the current arrangement more directly than the companies’ policy preferences did.

Henry de Zoete, AISI’s director, acknowledged that the institute relies entirely on cooperation from companies and foreign governments to obtain access. Cat Little, the senior Cabinet Office civil servant, confirmed that the regime was voluntary.

Asked what would happen if AISI judged a model dangerous but its developer released it anyway, de Zoete acknowledged that the institute had no power to prevent the release.

“We’re not saying yes or no to any sort of model,” he said.

That establishes an important boundary around government evaluation. Testing can produce evidence about a model’s capabilities and risks without functioning as a licensing decision. A company’s participation in an evaluation should not be mistaken for government approval of the resulting product.

The testimony also identifies two different limitations. AISI cannot compel the cooperation on which its testing depends, and its findings do not amount to a binding release decision. Addressing the first would not automatically resolve the second.

A law could, for example, require developers to provide testing access while leaving deployment decisions with the companies. A different framework could attach consequences to specified findings. These are possible policy designs, not powers the hearing established AISI currently possesses.

This is why the phrase “mandatory safety rules” needs careful treatment. Without knowing what is mandatory, who can enforce it, and what follows from a failed evaluation, the phrase says little about how much control a government actually gains.

British Testing Access Also Depends on Washington

The hearing revealed another constraint: a company’s willingness to cooperate with UK evaluators may not be sufficient.

Duff Gordon said OpenAI provides models before deployment to the US counterpart of AISI and would continue providing predeployment access to the British institute, subject to what the US government permits.

Anthropic described a similar limitation. Booth-Smith said the company had previously been unable to provide UK AISI with full, unrestricted access to a model, attributing that issue to the US government. He added that AISI now had access.

These were company accounts of the access arrangements, not independently established explanations of the underlying restrictions. They nevertheless complicate the assumption that British evaluators can secure everything they need through bilateral agreements with developers.

There are two potential points of discretion: the developer’s decision to cooperate and the permissions governing what it can share internationally. A commitment from a lab may not resolve both.

OpenAI’s support for an international convention is relevant in that context. Cross-border arrangements could become an important part of making national testing regimes workable. But support for a convention does not establish agreement on its membership, obligations or enforcement, and the reported testimony supplies no completed design.

For UK policymakers, the immediate challenge is therefore broader than persuading companies to sign stronger pledges. It includes establishing whether access can be dependable when the models and the authorities governing them sit outside Britain.

The Policy Divide Is Clearer Than the Proposed Law

The hearing put a consequential disagreement on the parliamentary record. Anthropic endorsed binding rules, OpenAI backed statutory obligations beginning with incident reporting, and Google DeepMind and Meta declined to make the same commitment to a new framework.

It did not establish that the first two companies support every possible intervention, or that the other two reject all mandatory oversight. Nor does the testimony establish why the companies differ. Assigning competitive motives would require evidence beyond these statements.

The next useful comparison is between concrete obligations. Would developers have to report defined incidents? Could evaluators require timely access to relevant models? Would a serious finding trigger disclosure, additional safeguards or a delay?

Those questions turn a general preference for safety into an enforceable system. The UK officials’ admissions make the starting point clear: AISI’s ability to examine frontier models currently rests on cooperation, and its evaluations do not give it authority to stop deployment. Any new framework should be judged by which of those limits it changes, rather than by how many companies endorse regulation in principle.