OpenAI and Anthropic tell UK lawmakers to regulate them; Google DeepMind and Meta decline

Two of the four frontier AI labs told British lawmakers on October 13 to regulate them, and the other two declined. At an oral evidence session of the Business, Innovation, Science and Trade Committee (BIST) in London’s Portcullis House, OpenAI and Anthropic called for binding rules — with OpenAI putting mandatory incident reporting first — while Google DeepMind and Meta stopped short of endorsing new legislation. The committee’s running order carried its own weight: former researchers turned whistleblowers at 2pm, the four labs at 3pm, and government witnesses at 4pm, including Henry de Zoete, a director at the AI Security Institute (AISI), who conceded on the record that the institute cannot stop a company from releasing a model it deems dangerous.

Key points

• OpenAI’s Tom Duff Gordon, VP and head of EMEA policy, backed a statutory framework and an international convention: “there should be some elements that should be mandatory,” starting with incident reporting. Basing frontier risk “on voluntary commitments is not a good long-term plan,” he said, per Wired.
• Anthropic’s Liam Booth-Smith, head of EMEA policy: “The rules should be binding,” and the company has “consistently believed that voluntary commitments are not enough” (per Wired).
• Asked directly whether he supported a statutory framework, Google DeepMind’s Owen Larter answered “Focus on the standards is what we would be doing”; Meta’s Rob Sherman pointed to existing UK regulators, which are “not writing on a blank page.”

Background

The session moves the UK’s AI safety debate from the government’s side of the street to the labs’ witness bench. On September 22, BIST invited the four companies to appear, citing growing AI safety concerns. The UK’s frontier-model regime has been voluntary: AISI evaluates the strongest models on the condition that companies hand them over. On the government side, the Financial Times reported on October 12 that Whitehall is drafting a compulsory safety-reporting and monitoring plan — we covered that in the UK’s compulsory safety reporting plan. This hearing is the labs’ answer, delivered in one shared parliamentary record.

The facts

Per the committee’s event page, evidence on October 13 ran in three panels: former researchers at 2pm, the four labs at 3pm, and government officials at 4pm. Where the four labs landed:

Lab	Witness	Title	Position on new legislation
OpenAI	Tom Duff Gordon	VP and head of EMEA policy	Backs a statutory framework and a convention; incident reporting first
Anthropic	Liam Booth-Smith	Head of EMEA policy	“The rules should be binding”; voluntary pledges not enough
Google DeepMind	Owen Larter	Senior director	No endorsement of legislation: “focus on the standards”
Meta	Rob Sherman	VP of global policy	Existing regulators cover it; “not writing on a blank page”

The government witnesses pinned the regime down as voluntary. Wired reports that AISI director de Zoete told MPs: “What we’re not doing is kite marking. We’re not saying yes or no to any sort of model.” Asked whether AISI could stop a company releasing a model it deemed too dangerous, he acknowledged it cannot — noting that frontier builders “are either American or Chinese,” and “There’s only so much we, as the UK, can do in that kind of context.” Cat Little, the Cabinet Office’s permanent secretary (AISI’s parent department), confirmed the current regime is voluntary: “At the moment, yes, it is.”

Anthropic added one access caveat of its own: Booth-Smith said the company had been unable to give UK AISI full unrestricted access to Mythos 5.1 at the time — “a question for the US government,” he said — though AISI now has access (per Wired).

The whistleblower panel sharpened the case for statute. Daniel Kokotajlo, a former OpenAI researcher: “I don’t think the companies can be trusted to self-regulate.” Steven Adler, who left OpenAI in late 2024, said he was “not confident in the safety culture across the industry.” Geoffrey Irving, formerly AISI’s chief scientist, rejected the idea that labs are locked in a race governments will not interrupt and cannot exit: “Both of those are false. They could just stop.”

What others say

Both pro-regulation labs gave their reasons in concrete terms. Per PA wire reporting carried by the Stourbridge News, Duff Gordon defended OpenAI’s decision to pull its latest model: “We found it didn’t refuse enough, it hadn’t progressed enough with alignment,” he said of GPT-6.1 Astra. “It is evidence that OpenAI takes safety really seriously — we are not racing to the bottom.” Booth-Smith put Anthropic’s exposure on the map: deployed models he considers safe, but “the surface area of immediate risk is definitely in the cyber domain” — Mythos has been finding tens of thousands of vulnerabilities a month — and “it is possible that real harm could be caused” through assistance with biological or chemical weapons.

MLex, the regulatory legal outlet, read the session as four companies that “defended different approaches” under questioning, with two points of common ground: support for external scrutiny and safety testing, and a willingness in principle to give AISI access to unreleased models — as long as they are “legally allowed” to. Committee chair Liam Byrne told Wired after the hearing that the session aimed to “light up the wide differences in opinions between some of the wealthiest firms on Earth” and show “that you couldn’t rely on voluntary self-regulation.” His proposal: use the UK’s upcoming G20 presidency (South Korea follows, per Wired) to broker an international AI convention independent of the White House — “The G20 presidency will outlive the US presidency of Trump.” Wired also notes that MI5’s director general Ken McCallum used his annual threat update the same day, a few hundred meters from the hearing room, to warn that AI could change “the stakes on existing or even new bioweapons.”

Our take

The split matters because of where the line falls: on whether there should be law at all, not on which rules. For companies building on these APIs, mandatory incident reporting is the obligation most likely to land first — OpenAI itself ranked it first, and the government’s draft plan starts there. Set the UK next to the American route — an executive order plus voluntary pledges, with enforcement arriving case by case (see the “superintelligence” order and regulators closing in on AI labs) — and legislation would put obligations before incidents rather than after them.

The honest catches: parliament has not yet published the oral-evidence transcript, so the quotes here rest on Wired’s and PA’s reporting; MLex sits behind a paywall and we verified only its headline and standfirst; and the FT’s compulsory-reporting plan remains a single-source story with no bill text. DeepMind’s and Meta’s positions survive in one line each in Wired’s account — worth re-checking against the transcript when it lands.

What to watch

• The transcript: parliament usually publishes oral evidence within days; it will settle what each lab actually committed to.
• The committee’s report: Byrne says BIST will spend significant time on the UK’s path as an AI “middle power.”
• The G20 convention idea, and whether DeepMind and Meta shift as the draft plan moves toward legislation.