Hassabis Wants a FINRA for Frontier AI — Industry-Funded, Federally Overseen, and Running Before Year's End

Demis Hassabis has spent a decade insisting that the people building artificial general intelligence should not be the only ones deciding when it ships. On Tuesday he proposed how to formalize that — and handed his own industry the bill.

In a manifesto published July 14 on his Substack and an X post since viewed some 8 million times, the Google DeepMind co-founder and CEO called for the U.S. to establish a Frontier AI Standards Body: an industry-funded, federally overseen organization with the staff and compute to test frontier models before release — and, eventually, the power to keep them off the U.S. market if they fail.

AGI is "probably only a few short years away," Hassabis writes; we are "standing in the foothills of the singularity." A few hundred words later he is discussing benchmark deprecation schedules. The model he reaches for is FINRA.

What he actually proposed

Hassabis's framing is narrower than the "independent regulator" shorthand that greeted it. He describes "a new Standards Body modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority (FINRA), with a board that includes independent leading technical experts and open-source representatives" — a majority-independent board, he told Axios, stacked with Turing Award winners. Funding "would need to be substantial and likely mostly come from industry."

The mechanics run in two phases. First, "Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release." Then the ratchet: "Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market."

The body would define "Frontier-class" via benchmarks it refreshes "perhaps quarterly to start," working with federal agencies and the U.S. National Labs to probe cybersecurity, biological threat uplift, and agentic deception. It would cover models "no matter their country of origin or whether they are open or closed," while exempting startups and academia. Its most aggressive lever arrives almost in passing: the regime "could be ratcheted up if the seriousness of the situation demands, including coordinating a slowdown in development among the Frontier Labs if deemed necessary."

Hassabis told Axios he spent months briefing the Trump administration, rival lab leaders, and European officials before going public. "The noises I've been hearing are very positive," he said. His timeline: "Months," ideally operational "before year-end." He called last month's improvised crackdown on Anthropic's Mythos and Fable models — frozen overnight by an export-control order, then negotiated back over two and a half weeks with no playbook — "a bit of a wake-up call."

What FINRA is, and what the analogy carries

FINRA is not a government agency. It is a private, industry-funded nonprofit that writes and enforces the rules governing U.S. broker-dealers, licenses securities professionals, surveils markets for manipulation, and arbitrates disputes — under SEC oversight, which must approve its rules. Member firms pay for it. It can fine, suspend, and expel.

Ported into AI, that structure delivers three things at once: delegated public authority, private funding from the regulated, and enforcement teeth. Nothing in AI governance has all three. The U.S. Center for AI Standards and Innovation and the UK AI Security Institute test models but cannot block releases. The EU AI Office has legal authority but sits outside the U.S. market. The Frontier Model Forum coordinates member labs but is voluntary all the way down.

That last gap is the one Hassabis targets, and he is not first to name it. Writing in Lawfare, Mark Thomas argued the same case under the headline "AI Companies Can't Regulate Themselves. They Should Regulate Each Other." Every major frontier lab except xAI already belongs to the Forum; what it lacks, Thomas wrote, is "statutory authority, mandatory membership, and government oversight."

Analysis: speed versus capture

The case for is partly political arithmetic. White House AI advisor Sriram Krishnan recently discounted any in-house AI regulator, telling the Financial Times "there will not be an FDA for AI." An SRO routes around that. It is also a talent argument: only industry money buys Turing-caliber evaluators and test compute, and a body rewriting benchmarks quarterly can track capability jumps no statute would survive. Hassabis's pitch is that the approach "would be technically focused, while at the same time supporting innovation and incentivising responsible behaviour."

Industry reception was warm, itself a data point. Satya Nadella called it "an important piece from Demis. We need more of this kind of thinking." Sam Altman called it "a thoughtful proposal." Even Elon Musk, whose xAI would be pulled into a body it never joined, allowed it was "a good starting point for discussions."

The case against starts with that unanimity. Critics argue the design invites regulatory capture on its face: the companies underwriting the body would help author the standards they are graded against, and the essay offers no structural fix — benchmarks would "initially" be set in consultation with Frontier Labs, with independent held-out tests arriving "eventually." FINRA's own record is the second problem; the same funding arrangement has long drawn criticism that it is a toothless insiders' club. And the conflict deserves naming plainly: a lab CEO is proposing the rules that would govern his lab, on a timeline he set, with a designation he concedes doubles as marketing — "a pretty nice, prestige kind of asset to have," he told Axios. His partial defense is the exemption for startups and academia, blunting the charge that incumbents are pulling up the ladder.

There is also a live dispute over who holds the gun. Anthropic CEO Dario Amodei wants an FAA-style federal agency with direct power to block unsafe deployments. The chiefs behind Gemini and Claude now agree Washington should regulate them; they disagree on whether the regulator should be the government.

What to watch

Whether the administration converts "very positive noises" into anything structural: an SRO needs a statutory hook and a supervising agency, and no bill exists. Whether rival labs back this in writing rather than in quote-tweets — Musk's support for a body xAI would have to join is the one to test. And whether a year-end deadline survives Washington, where institutions of this shape take years to stand up.

---

Sources: TechCrunch · Axios · Demis Hassabis, "A Framework for Frontier AI and the Dawning of a New Age" · Constellation Research · Lawfare