Everyone now agrees the United States should have some kind of referee for the most powerful AI systems. The interesting fight, suddenly, is over what it should look like, and this month that fight moved from theory to blueprint. On July 14, Google DeepMind chief executive Demis Hassabis published a framework calling for a US-led Frontier AI Standards Body modelled on FINRA, the industry-funded self-regulator that oversees stockbrokers under the Securities and Exchange Commission. Three days later, Bloomberg reported the White House was reviewing a proposal built on exactly that template, developed with involvement from Treasury Secretary Scott Bessent.
The appeal of the FINRA model is that the hard institutional question looks already answered. Under the proposal, frontier labs would submit models for review up to thirty days before release, during which the systems would be tested for dangerous cyber, biological and deceptive capabilities. Participation would start voluntary and harden into a condition of US deployment once the process proved itself. A body of that shape, reporting to an existing regulator, is a reasonable answer to a real coordination problem.
A trio of former government AI officials, writing for the Council on Foreign Relations, argue the institutional form was never the hard part. The hard part is five unresolved design questions, and the answers will decide whether the body earns decades of trust or has to be rebuilt after its first crisis. Independence comes first: a body funded by the companies it judges echoes the issuer-pays credit-rating model that helped detonate confidence in 2008. Then national security, because an evaluator holding pre-release access to every frontier lab instantly becomes one of the world's highest-value espionage targets. Then legitimacy abroad, guaranteed access to the models themselves, and the uncomfortable fact that there is no settled science for measuring frontier capabilities in the first place.
Their prescription is to separate the money from the judgment. The standards that define what "safe" means should be funded only by neutral, non-industry dollars, while industry money is confined to building shared testing infrastructure. Funding should be bound to guaranteed model access so neither can be quietly withdrawn. And the classified interface should be designed before the first classified finding, not after, because national-security designations expand to fill whatever space you leave undefined.
Money is already moving faster than the design. This week Anthropic announced a second $20 million donation to Public First Action, a bipartisan advocacy group, doubling its total to $40 million; chief executive Dario Amodei has personally given $1 million to an allied super PAC. Anthropic wants government to hold the power to block the release of a model that fails safety standards, plus mandatory third-party audits. Rival executives, bankrolling their own super PAC called Leading the Future, want lighter rules so the US can outrun China. The referee has not been built yet, and the teams are already lobbying over how to pick the umpire.
Lurking behind the whole debate is a more radical idea that both parties, unusually, seem to like: not regulating the labs but owning them. A Vermont senator has floated requiring the big labs to hand half their equity to a sovereign wealth fund, and the president has mused that a government stake "would be a beautiful thing." Governor Gavin Newsom has separately proposed a national public equity fund so Americans can own a piece of the industry. It is the difference between hiring a referee and buying the team, and the fact that it is on the table at all shows how unsettled the question of public control over AI still is.