AI firms split on binding UK rules as officials spell out AISI’s limits

Anthropic, OpenAI, Google DeepMind and Meta offered different answers on binding AI-safety rules at a UK parliamentary hearing, according to NeoTeo’s account. The practical point: officials described an AI Security Institute that could evaluate models, but neither certify them nor stop their release. That separates two questions often bundled together: how Britain tests powerful AI, and what authority follows from those tests.

UK AI Security Institute Watch analysis

What happened

In its report published on 14 October, NeoTeo summarised coverage by WIRED and Infobae of the companies’ testimony. Anthropic’s head of Europe, Liam Booth-Smith, supported binding rules. OpenAI vice president Tom Duff Gordon supported a statutory framework with some mandatory requirements, starting with incident reporting.

Google DeepMind’s Owen Larter emphasised standards, while Meta’s Rob Sherman pointed to existing UK regulators. Those positions are not interchangeable, and the reported exchange does not establish that either company opposed every possible new requirement.

Officials described the regime at the time of the hearing as voluntary. AISI director Henry de Zoete said access to leading models depended on cooperation from companies and foreign governments, and that the institute did not approve or reject releases.

Why it matters

Testing capability and regulatory authority are different things. An institute can investigate a model’s behaviour without having the power to require access, attach conditions to deployment or prevent a launch. Readers should not mistake an AISI evaluation for a government safety certificate.

The companies’ answers also reveal different starting points for regulation: binding safety duties, selected mandatory measures, technical standards and existing regulators. These could overlap in a future framework, but they distribute responsibility differently. Mandatory incident reporting, for example, would create a specific disclosure obligation rather than leave reporting entirely to company discretion.

Our read

The useful story is not a neat contest between companies that favour safety and companies that do not. It is a disagreement about which obligations should be enforceable, and who should enforce them.

Britain’s next policy test is whether it wants AISI to remain principally an evaluator or become part of a system with compulsory duties. Calling something a watchdog does not, by itself, supply the teeth.

For now, distinguish testimony from legislation. The hearing records positions and institutional limits; it does not announce new powers or a binding settlement.

What to watch

• Whether the committee recommends mandatory model access, incident reporting or other specific duties.
• Whether government proposals change AISI’s authority rather than simply expand its testing work.
• How the companies translate their broad positions into support for, or objections to, concrete legislative provisions.

Discussion spark: Should the UK give AISI compulsory model-access and release-blocking powers, or keep it as an evaluator while existing regulators enforce narrower duties such as incident reporting?

not affiliated with or endorsed by UK AI Security Institute