LONDON – Demis Hassabis, the chief executive of Google DeepMind, proposed on Monday a financial-industry-style regulatory framework for artificial intelligence: a standards body modeled on FINRA that would review frontier AI models before their release, staffed by technical experts and funded by the labs building the systems it assesses.
The proposal responds directly to criticism of recent ad hoc safety evaluations, including scrutiny of Anthropic’s Mythos model and OpenAI’s Sol system, which drew questions about the depth of technical review and the opacity of recommendations that labs were not required to follow. Under Hassabis’s plan, the body would hold a 30-day window to assess models before each release – short by most regulatory standards, but considerably longer than existing informal pre-release evaluations.
“The strength of this approach is it would be technically focused, while at the same time supporting innovation,” Hassabis wrote in a public post outlining the proposal, adding that frontier labs would initially share models with the body up to 30 days before release on a voluntary basis. The body would be backed by the United States government but financed by the AI industry and governed independently – similar to how FINRA operates under the Securities and Exchange Commission without being a government agency.
Staffing would draw from open-source communities and technical practitioners, and the structure is designed to reduce dependence on legislative action. Hassabis floated the possibility that what starts as voluntary participation could eventually become a de facto requirement for United States market deployment – a mechanism to enforce compliance without a congressional mandate. The body would also collaborate with labs after release to address critical vulnerabilities discovered in deployed systems.
The White House has already ruled out the more prescriptive alternative. Sriram Krishnan, the administration’s AI policy adviser, said recently that there will not be an FDA for artificial intelligence – rejecting the pre-market approval model that some safety researchers have proposed. That position created a gap the DeepMind proposal is explicitly designed to fill: a formal review window, real technical expertise, and independent governance, without requiring a vote in Congress.
The frontier model market that Hassabis is trying to bring under a governance framework is moving quickly. DeepSeek is in talks to raise $1.5 billion at a $71 billion valuation ahead of a planned IPO, with Chinese state investors and Tencent among the backers. A self-regulatory body covering only U.S.-based labs or only those that choose to participate would face immediate questions about whether it actually addresses the conditions that make frontier model releases high-stakes in the first place.
The proposal’s most significant structural limitation is also its most explicit one. A lab that submits a model, receives critical findings, and launches anyway faces no enforcement mechanism under the body as described. Hassabis acknowledged the voluntary nature of the initial framework, but the path to mandatory compliance he sketched depends on an authority that the body has not yet earned – and that no existing institution has so far been willing to confer.
That question matters differently depending on which lab you’re talking about. DeepMind’s proposal would apply to its own systems as much as to any competitor’s, which gives it a credibility that purely industry-association governance frameworks lack. Hassabis is proposing a structure that constrains himself along with everyone else. Whether that gesture converts into a functioning body depends on whether OpenAI, Anthropic, and other frontier labs agree to participate – and whether participation under a 30-day pre-release window produces reviews meaningful enough to delay a launch when the findings warrant it.
The FINRA parallel Hassabis invokes is instructive in ways that complicate the comparison. FINRA can impose fines, suspend licenses, and bar individuals from the industry. An AI standards body built on voluntary industry funding and optional submission would have none of those mechanisms at launch. The credibility of any findings it produces, and the pressure on labs to take those findings seriously, would depend on whether the reviews are technically rigorous enough to be seen as authoritative rather than convenient. That is precisely what existing evaluations have been criticized for failing to achieve.
