Hassabis pitched a Wall Street-style AI watchdog to Bessent and Kratsios. The Treasury has its own version.

Demis Hassabis, the co-founder of Google DeepMind, met with Treasury Secretary Scott Bessent and White House science policy director Michael Kratsios this week to pitch a private, industry-funded AI safety body modeled on FINRA, the private watchdog that polices Wall Street under SEC oversight. The meeting, first reported by the Wall Street Journal on August 12, lands the same week the Treasury is building its own separate architecture for governing advanced AI, surfacing two parallel governance tracks at the moment the country is about to pick a default for the most powerful AI systems.

The proposal, which Hassabis has been developing publicly for nearly a month, would create a self-regulatory organization (SRO) for frontier AI, the most capable, general-purpose models from leading labs. Frontier labs would initially submit new models voluntarily up to 30 days before release, with mandatory U.S. deployment tests only after the protocol proves robust. The body could, in principle, coordinate an industry-wide slowdown if risks mount. It currently has no government approval or binding authority. Hassabis published the governance manifesto on Substack on July 14, calling the model "a FINRA for frontier AI" and comparing its scale of ambition to the International Atomic Energy Agency.

The Treasury's separate effort, first reported in July, would create an independent federal agency to oversee frontier AI models directly, a public-cop alternative to Hassabis's industry-funded design. The two tracks have not yet publicly collided. They are now being shopped at the same level of government, within weeks of each other, with overlapping personnel and overlapping jurisdictions.

A voluntary federal regulatory framework for advanced AI models was finalized on August 4, eight days before the Hassabis meeting became public. That framework sets the baseline any new body would have to coexist with. The window between the framework's publication and Hassabis's August sit-downs suggests the administration is now entertaining architectures on top of it, not instead of it.

Hassabis has also raised the proposal with executives at major AI laboratories to build political and industry consensus in parallel. His target date for an operational body is the end of 2026, a tight clock given that any SRO still needs a charter, funding, and at least one federal agency willing to host it. A CFR analysis of the pending U.S. AI regulator lays out the design questions: who funds it, who staffs it, what its enforcement teeth are, and how it coordinates with existing agencies.

The meeting was separately confirmed by Bloomberg on July 16 and by Briefs.co.

The contest now visible is structural. A FINRA-style body would be funded and staffed by the labs it regulates, with a federal overseer holding a veto, much as FINRA member firms today underwrite the regulator that polices broker-dealers. A Treasury-led independent agency would be a public regulator with statutory authority. Both architectures are serious; both are also untested at frontier scale. The asymmetry is funding. A private SRO paid by the largest labs has every incentive to set thresholds the largest labs can already meet, while a public agency can hold the bar at the level the public is willing to fund.

The next tests are concrete. Treasury's separate track has to produce a written proposal in the same window. The voluntary 30-day submission window in Hassabis's plan has to attract real sign-ups before either architecture hardens into a default. The public will be living with whichever body wins for at least a decade.