Hassabis proposes an AI FINRA - 30 days before each release

In plain terms

The CEO of Google DeepMind proposes creating an independent body to evaluate the most advanced AI models before they are released to the market. The model he invokes is that of the American FINRA: a structure funded by the sector it controls, but which operates autonomously.

Ce qui est proposé, exactement

Demis Hassabis detailed the mechanism on July 14, 2026. The device would be progressive: initially, voluntary participation of laboratories, which would share their models with the body up to 30 days before their release. His wording: "Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release." Once the process is tested, the evaluation would become mandatory for any deployment in the United States. The body would also manage critical vulnerabilities discovered after launch.

The context matters: the proposal comes after the criticisms addressed to the ad hoc government reviews of the Mythos (Anthropic) and Sol (OpenAI) models, deemed weak technically and opaque.

Under the hood

Delegated self-regulation is neither new nor an aberration. FINRA exists because the public regulator has never had the means or expertise to monitor tens of thousands of brokers in real time. It is funded by the entities it controls, has the power to sanction, and operates under the supervision of the SEC. This last point is decisive: self-regulation only works if a public regulator can disavow it.

Transposed to AI, the reasoning holds. Expertise in evaluating frontier models is counted in a few hundred people worldwide, almost all employed by the laboratories. An independent body funded by the industry may be the only way to pay these people at market prices without waiting for a state to vote a budget.

Les angles morts

Three, which must be named without intent.

The capture. Who appoints the council? Who arbitrates a disagreement between the body and a laboratory that finances it? Without an equivalent public oversight to that of the SEC, "independent" is an adjective, not a structure.

The 30 days. The window is calibrated for a compliance audit, not for adverse research of flaws. The reminder came this week: IEEE Spectrum documented eight distinct jailbreak techniques that affect most of the commercial industry - found not in a pre-launch review, but afterwards, by researchers outside the laboratories. An evaluation that begins thirty days before release will look for what it has been taught to look for. The flaws that matter are those that no one has thought of yet.

The national perimeter. An obligation limited to U.S. deployments leaves intact the open models published elsewhere - and the figures from Hugging Face remind us that the Chinese share of actual usage is no longer marginal. We would regulate the most visible channel, not the widest.

So what

The proposal deserves better than the cynical reflex of the "fox guarding the henhouse." Faced with improvised government reviews, a technical device, funded, with a binding deadline, is a clear progress.

But the history of FINRA also teaches that real power is not in the body: it is in the public authority that can contradict it. As long as this authority is not named, what Hassabis proposes is a very good evaluation laboratory. Not a regulator.