Hassabis Wants a FINRA for AI, and So Does Washington

A FINRA for AI, proposed by the industry and echoed by the Treasury

The most consequential idea in AI governance this month is not a bill or an executive order. It is an institution that does not exist yet, and the notable thing about it is who is drawing the blueprint. Demis Hassabis, following his July framework on frontier AI, has been personally briefing Treasury Secretary Scott Bessent and White House science director Michael Kratsios on a proposal to create a standards body for frontier models built on the shape of FINRA, the industry-funded, government-overseen regulator of American brokerages.

The mechanism is specific. Frontier labs would submit new models roughly 30 days before release to a central body that tests them for the capabilities that keep policymakers awake: cyber offense, biological uplift, and deception. Participation would begin voluntary and become mandatory for deployment in the US market once the process proves it works. It is self-regulation with a government backstop, the same bargain that governs stockbrokers, applied to the most powerful technology of the decade.

Why the same design appeared on both sides of the table

What lifts this above the usual lobbying is that the government was already sketching the same institution. Bessent, according to reporting on the Treasury's internal work, had developed a comparable FINRA-model proposal now under review by the White House chief of staff. The industry's most prominent safety advocate and the administration's top economic official arrived at nearly the same design without coordinating on it.

That convergence is not a coincidence. It is what happens when every other option has been tried and found wanting. Hard statutory regulation moves too slowly for a field that ships a new frontier model every few weeks. Pure voluntary commitments have no teeth. And the improvised alternative, which Washington has actually been running, is worse than either.

He has a point that Santage readers have watched unfold. The government froze one lab's advanced models for more than two weeks under export-control authority, then negotiated a restricted launch arrangement with another. Those were decisions made case by case, behind closed doors, with no published standard anyone could plan around. A FINRA-style body is an attempt to replace that improvisation with a rulebook. The appeal, to a lab, is predictability. The appeal, to the Treasury, is leverage that does not require an act of Congress.

The capture problem the model cannot design away

The trouble sits inside the design. A body funded by the industry it polices, staffed by the only people with the expertise to police it, and shaped in private consultation with the executives it will regulate, is a textbook setup for regulatory capture. The Council on Foreign Relations, examining the proposal, flagged five unresolved problems, and capture is the one that structure cannot easily solve: when the regulated write the rules, the rules tend to protect the regulated.

A referee that cannot reliably measure fouls, funded by the players, is not obviously better than no referee at all. That is the gap between the appeal of the idea and the difficulty of building it well.

What voluntary-then-mandatory means for every frontier lab

For the labs, the strategic calculation is already clear enough that most are cooperating. Sam Altman called the framework thoughtful, Elon Musk called it a reasonable starting point, and even reliable regulation skeptics conceded it had merit. That unanimity is itself a tell. When every major lab endorses a proposed regulator, the regulator is probably one they expect to be able to live with, and possibly one they expect to shape.

The deeper shift is that voluntary is a waypoint, not a destination. The design's whole logic is that once the testing process is established, compliance becomes the price of selling a model in the United States. A lab that helps build the body now is buying a seat at the table when the voluntary phase ends and the mandatory one begins. A lab that stays out risks having the standard written without it, then applied to it anyway.

The question Washington has not answered is who the body ultimately serves. Built well, an industry-funded standards regulator could give the public a real check on frontier capability without freezing innovation, the way FINRA polices brokers without shutting markets. Built the way it is currently being drafted, in private conversations between the regulator's future subjects and the officials who will authorize it, it could just as easily become a compliance ritual that launders each new model as safe. The design is converging. Whether it converges on oversight or on theater is the fight that has not started yet.