DeepMind's Demis Hassabis Pushes for an Independent AI Safety Oversight Body

# DeepMind's Demis Hassabis Pushes for an Independent AI Safety Oversight Body

Demis Hassabis has stopped writing memos about how to govern advanced artificial intelligence and started lobbying the people who could actually build the machinery. The Google DeepMind co-founder, who stepped back from running the lab this month to become its chairman and Alphabet's chief scientist, has been privately pitching senior U.S. officials and rival lab executives on an independent body to set common safety practices for the industry as it races toward more capable systems, according to a Wall Street Journal report. Among those he has approached: Treasury Secretary Scott Bessent and Michael Kratsios, the White House science adviser who directs the Office of Science and Technology Policy.

The pitch marks the moment a debate that has largely played out in op-eds and conference panels crossed into active political negotiation. Hassabis is not merely floating an abstraction. He wants a working institution, and he has said he wants it operational before the end of 2026.

From manifesto to meeting rooms

The blueprint went public on July 14, when Hassabis released an essay titled "A Framework for Frontier AI and the Dawning of a New Age." It laid out an organization modeled on the Financial Industry Regulatory Authority, or FINRA, the industry-funded body that polices Wall Street brokers under the Securities and Exchange Commission's oversight. The proposed Frontier AI Standards Body would be funded by the AI industry, staffed by top technical experts, and answerable to the U.S. government without being a government agency itself.

The mechanics are gradual by design. Frontier labs would initially share their most capable models with the body voluntarily, up to 30 days before release, for testing that probes dangerous cyber, biological and "deception" capabilities. Only once the regime proved effective and robust would passing become mandatory for deployment in the U.S. market. Hassabis has also floated a more dramatic power: the body could coordinate an industry-wide slowdown if risks mounted. He has told colleagues he believes severe capabilities could reach openly available models within roughly 18 months, which lends his year-end deadline its urgency.

What changed in August is the venue. According to the Journal, Hassabis raised the concept not only with Bessent and Kratsios but with executives at other major AI labs, a deliberate effort to build political and industry consensus before any structure is locked in. He has framed the current improvised approach as untenable. When the Trump administration briefly held up one of Anthropic's most advanced models this summer under an export-control directive, Hassabis told Axios it was "a bit of a wake-up call," evidence that Washington needs a formal governance structure rather than case-by-case interventions with no transparent process.

The IAEA analogy, and where it breaks

Hassabis and others have reached for grand comparisons to convey the stakes, invoking the International Atomic Energy Agency as a model for a global body that certifies standards and monitors compliance. The analogy captures the ambition: a shared, credible referee for a technology whose worst outcomes could be catastrophic and cross-border.

But the parallel strains quickly. The IAEA governs a physical input, enriched uranium, that is scarce, trackable, and hard to produce. Frontier AI models are software. A pre-release review regime anchored in the United States can impose real compliance costs on American and allied labs, but it has no obvious mechanism to force a Chinese developer such as Moonshot AI, which Kratsios publicly accused in July of illicitly obtaining Nvidia chips, to submit a downloadable model for inspection before releasing it. Hassabis's framework asserts that standards should apply to all frontier-class models regardless of origin; it does not explain how they would be enforced beyond U.S. jurisdiction. That enforcement gap is the ceiling every version of this idea keeps bumping against.

Self-regulation, or a template for capture?

The more revealing distinction is not IAEA versus FINRA but industry-run versus government-run. Hassabis has deliberately positioned his proposal between two rivals. OpenAI's Sam Altman has pushed something closer to an IAEA-style international forum that uses market access as leverage. Anthropic's Dario Amodei has favored an FAA-style agency with direct blocking authority vested in the government, an option the Trump White House has explicitly ruled out. Hassabis's industry-funded, government-overseen middle path has, so far, threaded a needle the others have not.

The endorsements are strikingly broad for a regulatory question. Altman called the proposal "thoughtful." Elon Musk described it as "a thoughtful framework overall and certainly a good starting point for discussions." Even David Sacks, the former Trump AI czar who reflexively opposes licensing regimes, allowed that it had merit and beat the alternative of the government regulating frontier models directly. That an idea can win nods from Altman, Musk and Sacks at once is a testament to Hassabis's positioning, though skeptics would note it may also indicate a proposal the labs find comfortably survivable.

That is the core tension. A self-regulatory organization is, by definition, funded and staffed by the industry it polices. The Council on Foreign Relations, in a July analysis led by former NSA chief AI officer Vinh Nguyen, laid out five unresolved problems: independence, national security, talent, measurement, and legitimacy. The independence worry is the sharpest. The closest analogy to companies paying their own evaluators is the issuer-pays credit-rating model that helped corrode trust in 2008. And the expertise needed to stress-test frontier models sits almost entirely inside the very labs that would be reviewed, guaranteeing a revolving door.

Whether rival labs would genuinely accept binding oversight is the open question underneath the warm quotes. Voluntary 30-day reviews cost little. Mandatory market-access gates that a competitor's model could fail are a different proposition, and each lab's enthusiasm tracks its competitive interest. Analysts have noted, pointedly, that a referee able to slow the race is most attractive to the lab that feels it is losing.

What to watch

The most consequential wrinkle is that Hassabis is not lobbying a blank slate. Bloomberg has reported that Bessent has been developing a comparable FINRA-style concept inside the administration, a proposal that reached White House Chief of Staff Susie Wiles for review. Two parallel tracks are now taking shape: one from the private sector, one from within government, converging in private meetings the public learns about one report at a time. If the institution that emerges is designed largely in conversations between the industry's most influential figure and the Treasury Secretary drafting the template, the independence it claims may be compromised before it opens its doors.

Watch three things. First, whether the White House's separate voluntary review framework, built under Trump's June executive order and explicitly barred from becoming a licensing regime, hardens into something with teeth. Second, whether any binding commitment materializes from labs other than DeepMind, the real test of consensus. Third, whether the China enforcement gap forces the proposal to shrink into a purely domestic, allied-markets standard. The defining AI policy question for the rest of 2026 may be whether these tracks fuse into one durable body or splinter into competing frameworks that make governance harder, not easier.