A split, on the record

Select committees usually extract careful, convergent answers. This one extracted a disagreement, and appears to have been designed to. Four frontier labs gave evidence on Tuesday, and on the central question — should AI safety duties be law? — they divided two against two.

Liam Booth-Smith (Lord Booth-Smith), Anthropic's head of Europe, Middle East and Africa policy, was the bluntest: "The rules should be binding." OpenAI's Tom Duff Gordon was more measured but landed in a similar place, backing a statutory framework and an international convention, with mandatory incident reporting as the starting element. His reasoning, as reported, was that frontier risk is a poor thing to rest entirely on voluntary commitments over the long term.

Google DeepMind's Owen Larter, asked the same question directly, redirected to standards. Meta's Rob Sherman redirected to institutions, noting that British regulators are "not writing on a blank page." Neither answer is a refusal to be scrutinised. Both are refusals to ask for a statute, and the distinction is the story. The disagreement is over whether existing standards and regulators are sufficient, or whether frontier AI development needs additional binding obligations.

The promise with a condition attached

Both companies indicated they would respond if the AI Security Institute or British intelligence agencies presented evidence that a model was unsafe. Their answers covered possible steps such as pausing deployment or withdrawing a model, but those actions should not be treated as one identical, unconditional commitment. OpenAI acknowledged that it cannot necessarily withdraw deployments controlled by third parties.

Then came the qualifier, from OpenAI, and it deserves equal billing. A lab can only pull what it controls. Models hosted by someone else sit partly outside that reach. That is not evasion; it is an accurate description of how frontier models are distributed. It is also a precise map of where a voluntary kill switch stops working, and therefore an argument, made inadvertently by the firms themselves, for rules that attach to deployment rather than to the developer's goodwill.

The same pattern showed up in the access discussion. MLex reports that all four companies broadly committed to giving the Institute unreleased models — as long as they are legally allowed to. Readers of this site have seen that clause operate in practice: last month, pre-release access for a major model stopped at the US border, and was later restored from Washington. A commitment with a legal-allowance clause is a commitment whose real terms are set somewhere else.

What to watch, practically

The hearing changed no law. Two labs endorsed binding requirements, while two declined to commit to a new statutory framework. The UK AI Security Institute currently relies on voluntary arrangements for access to frontier models and has no independent statutory power to veto their release. The next questions are whether incident reporting becomes mandatory and whether pre-release testing, access and intervention acquire enforceable legal backing.

For anyone building on these models in the UK, the concrete thread to follow is incident reporting: it is the one mandatory element the pro-legislation side named first, it is the easiest to draft, and it would be the first duty to arrive if the government moves. The second thread is whether the Institute's access gets a statutory footing, because Tuesday's testimony described, in the witnesses' own qualifications, why voluntary access keeps developing exceptions.

Sources

  1. [1] UK Parliament — October 13 evidence session record
  2. [2] MI5 — October 13 annual threat update (published speech)
  3. [3] WIRED, "OpenAI and Anthropic Tell British Lawmakers: Regulate Us," 13 October 2026 (read via syndication)Read source
  4. [4] MLex (Patricia Figueiredo), "Pressed by UK lawmakers, Anthropic, Google, Meta, OpenAI differ on AI regulation," 13 October 2026Read source
  5. [5] Sky News via Capital, "Tech firms promise to take down AI platforms if British secret services show they are unsafe," 13 October 2026Read source