AI Regulation: OpenAI and Anthropic Back UK Rules
OpenAI and Anthropic backed binding AI rules before UK lawmakers. See what the hearing reveals about safety, access, and the future of AI governance in Britain.
OpenAI and Anthropic Want AI Rules. Their Rivals Don't.
The question of who should rein in advanced AI is no longer just something policymakers argue about in public. At a UK parliamentary hearing, representatives of OpenAI and Anthropic said voluntary safety commitments don't go far enough - and that some rules should actually be enforceable. Google DeepMind and Meta pushed back, leaning on standards and the UK's existing regulators instead of backing new legislation.
It matters because the UK's current setup leans heavily on cooperation. The AI Security Institute tests frontier models only when companies provide access. MPs were also told the government can't stop a firm from releasing a model the institute thinks is dangerous. And MI5 chief Ken McCallum warned that AI could raise the danger from existing or even new bioweapons.
The hearing lays out the practical dilemma for governments, developers, and businesses: how do you oversee systems that can create public risks when the decision to give access - and the decision to release - still sits largely in private hands? Here's what the companies said, where the UK system falls short, and what their disagreement could mean beyond Britain.
The hearing exposed a clear divide
The Business, Innovation, Science and Trade Committee brought four frontier AI companies, four former employees who had turned into whistleblowers, and senior civil servants into a session with a cross-party group of MPs. On the record, the core disagreement is usually framed simply: should oversight rely on company promises, or should governments require certain actions through binding rules?
Anthropic's head of Europe, Liam Booth-Smith, made the case for binding requirements. He told MPs that the company had long believed voluntary commitments aren't sufficient.
OpenAI vice president Tom Duff Gordon sounded less absolute. He backed a statutory framework and an international convention, and he said mandatory incident reporting should be a starting point. But he wasn't arguing that everything in frontier AI development should immediately be controlled by law - more that leaving frontier risks to voluntary commitments alone isn't a solid long-term approach.
Google DeepMind and Meta didn't back that approach. Owen Larter from Google DeepMind emphasized standards, while Meta's Rob Sherman pointed to existing UK regulators, arguing the country isn't starting from zero. The difference between them isn't just about how worried they sound - it's about whether new regulation is needed, and if so, what form it should take.
Why the disagreement matters
Voluntary commitments can move faster than legislation. A company can change its internal safety process or agree to share test results without waiting for a bill to pass. Standards can also help by giving companies and regulators a shared technical reference point - especially when model capabilities are changing quickly.
But voluntary commitments are still promises made by the very organization being assessed. They can be revised, interpreted narrowly, or become harder to enforce when business incentives shift. Laws can spell out minimum duties and consequences for failing to meet them. That doesn't guarantee good outcomes - laws can be poorly designed or poorly enforced - but it does make the obligations less dependent on a company's willingness to cooperate.
Committee chair Liam Byrne said the hearing's goal was to surface how the major AI firms disagree - and to push policymakers away from assuming self-regulation alone will cover the problem. The session didn't produce a single winner. It did make one thing clear: the industry isn't speaking with one voice, and regulators will have to plan around that reality rather than pretending otherwise.
Why voluntary oversight has limits
Some of the sharpest evidence came from officials describing what the AI Security Institute can do - and what it can't. The institute tests frontier models before release, but it relies on companies choosing to provide access. Henry de Zoete, the institute's director, told MPs it also depends on cooperation from foreign governments to get access.
Cat Little, a senior civil servant at the Cabinet Office, was asked whether the regime is voluntary. Her answer was blunt: at that point, yes. De Zoete also acknowledged the institute can't stop a release if it judges a model too dangerous but the company releases it anyway. The institute doesn't issue a safety certification that decides whether a model can go to market.
That creates a mismatch between assessment and authority. A government body can build expertise, run tests, and identify risks - yet still lack the information or legal power needed to make its findings matter in practice. As officials described it, the system depends on companies both providing access and taking the resulting concerns seriously.
Testing is not the same as control
Testing can be useful even if it isn't a licensing system. An institute can point to weaknesses, help improve evaluation methods, and give policymakers a clearer picture of emerging capabilities. But those roles shouldn't be treated as the same thing as the power to require testing, delay a release, or mandate specific safeguards.
This distinction matters for businesses that build products on top of frontier models. A developer might make product, investment, or security decisions using a provider's published policies or assurances. But if the underlying model hasn't been assessed independently - or if the assessment is confidential and access depends on the provider - then the downstream business has little ability to verify what it's relying on.
The hearing didn't argue that every model should require government approval before release. That would be a much bigger policy change with real costs for competition, research, and how fast products can ship. The narrower question was more operational: what happens when a public assessment finds a serious risk and the developer disagrees - or decides to proceed anyway?
Without a defined process for that scenario, testing risks becoming advisory by default. And the problem isn't that companies are automatically reckless. It's structural: the organization being assessed controls access and release, while the public body has limited leverage. That may be workable for lower-risk systems, but it's a shaky basis for managing risks officials say could be severe.
The UK's leverage - and its constraints
The UK can build technical expertise and set expectations, but it doesn't control every frontier developer. De Zoete noted that key companies are based in the US or China, and that there are limits to what Britain can do on its own. The hearing also showed that access to models can depend on decisions made outside the UK.
Duff Gordon said OpenAI provides predeployment access to CAISI, its US counterpart, and would keep providing models to the UK institute as long as it can do so under US government rules. Booth-Smith said Anthropic had previously been unable to provide full unrestricted access to a model discussed in the hearing, Mythos 5.1, because the issue was with the US government. He said the UK institute later gained access.
The takeaway is straightforward: the UK's ability to inspect some systems is shaped by companies' home jurisdictions and by government decisions there. A national institute can be capable, but capability doesn't guarantee access. If the information needed to assess risk is controlled elsewhere, national oversight has a built-in ceiling.
International rules are difficult, but the alternative is fragmentation
Byrne argued that the UK should use its coming G20 presidency - alongside South Korea's subsequent presidency - to pursue an international convention not dependent on the US political cycle. His reasoning was that the G20 timeline would outlast Donald Trump's US presidency.
An international agreement could help set shared expectations around things like incident reporting, access for evaluators, and how governments communicate serious risks. It could also reduce the chance that companies face incompatible requirements in different markets. Still, getting there is hard: countries differ in priorities, legal systems, strategic interests, and their relationships with the companies building these tools.
That's why international coordination can't replace domestic rules. An international convention can create common ground, but it doesn't automatically give a UK regulator the authority to require a test or act on its results. By contrast, national rules can create enforceable duties, but they may not reach companies operating overseas.
A workable approach likely needs more than one layer: domestic obligations that can be enforced, technical standards that can be updated, and international cooperation when risks and companies cross borders. Where those pieces balance out is a policy choice. The hearing suggests that leaving the balance unclear means leaning on access and restraint that governments don't fully control.
What the debate means for AI developers and businesses
For companies building with AI, this debate can sound abstract - but it affects delivery. Regulation and safety practices influence whether a model is available, what testing a provider must perform, how incidents are reported, and what evidence customers can expect when they're trying to judge risk.
If you build a product around an external model, you already face dependency questions: what happens when the provider changes access, restricts capabilities, or updates its terms? The UK hearing put a governance version of that question on the table: what happens if the model's safety hasn't been independently assessed - or if a public body can't share results with customers?
Companies don't have to wait for a final law to improve their own decisions. They can separate provider claims from evidence they can actually verify, document which model capabilities matter to the business, and plan for changes in access or acceptable use. These are operational controls, not a substitute for regulation.
A practical checklist for AI buyers
When choosing a model or AI provider, ask questions that tie governance directly to delivery:
- What has been tested? Ask what evaluations were done, by whom, and whether the results match your intended use case.
- What is the incident process? Confirm how the provider reports significant failures, vulnerabilities, or unexpected behavior.
- What happens when access changes? Identify a fallback plan if a model is withdrawn, restricted, or materially changed.
- Who owns the risk? Make responsibility clear across the model provider, integrator, and your organization.
- What evidence can you retain? Keep records of model versions, evaluations, approvals, and important provider communications.
This matters even more when AI supports a core workflow instead of a one-off experiment. A customer-support assistant and an AI system used for health, employment, or security decisions don't carry the same consequences. Your level of verification should reflect the impact of failure.
The whistleblower testimony is a governance signal
Four former researchers also testified. Former OpenAI staffer Daniel Kokotajlo said he didn't think companies could be trusted to self-regulate. Steven Adler, who left OpenAI in late 2024, said he wasn't confident in the industry's safety culture. Geoffrey Irving argued that companies' belief they're trapped in a race governments won't interrupt - and that they can't walk away - is mistaken.
These are testimonies, not proof that every company will behave the same way or that every risk is imminent. What makes them relevant is narrower: internal confidence and external accountability aren't interchangeable. A company can have skilled safety teams and still have incentives to prioritize release speed, market position, or customer demand. In other words, governance has to factor in incentives - not just assume good intentions will be enough.
For businesses, the point is practical. Don't treat a provider's safety policy as the whole of your risk management. Assess the model in your context, understand where the provider's commitments end, and decide based on the likely impact of failure - not only on how quickly a feature can go live.
Conclusion: Make accountability part of the system
The hearing didn't produce a final blueprint for AI regulation. But it did underline a basic point: the UK's oversight depends on voluntary access and cooperation, while some of the risks discussed by officials could spill well beyond the companies building these systems.
For policymakers, the question isn't simply regulation versus innovation. It's whether oversight should stay dependent on private cooperation, and what enforceable duties make sense for the most consequential systems. For businesses, the immediate job is to treat AI governance as part of product and operational risk - not something you revisit only when a new law passes.
The next step is practical: map where your organization relies on AI, identify what evidence supports those dependencies, and decide what you'll do if access, safety assumptions, or rules change. Read the original reporting by WIRED for the full account of the parliamentary hearing.