America Weighs an AI Watchdog the Labs Would Fund

A Wall Street model, pointed at AI

A plan taking shape in Washington would create an independent regulator to vet the most capable AI models, reporting to the Securities and Exchange Commission and modeled on FINRA, the private body that polices stockbrokers. Treasury Secretary Scott Bessent helped develop it, and it now sits with White House chief of staff Susie Wiles. The template is not new: Google DeepMind's Demis Hassabis published a governance proposal for exactly such a Frontier AI Standards Body weeks earlier.

The choice of FINRA as the model is the whole story. FINRA is private, funded by the industry it regulates, staffed largely from that industry, and supervised at arm's length by the SEC. Copied to AI, it means the frontier labs would fund and help run the body that inspects their own models before release.

Why it matters: voluntary now, mandatory later

The mechanism starts soft and hardens on a timer. In the first phase, labs would submit frontier models voluntarily for up to 30 days of review before release. The tests target dangerous capabilities rather than conduct: autonomous action, guardrail bypass, deception, offensive cyber ability, and biological or nuclear risk.

Once the assessment protocol is judged robust, passing it would become a condition of deploying a model in the US market. The voluntary opening is the on-ramp to a mandatory gate, reached without new legislation, which is precisely why it can move faster than a law fought through Congress.

The capture question worth asking

A body funded by the firms it inspects is faster to build but harder to trust. That is the tension FINRA already carries on Wall Street, where critics argue an industry-paid regulator pulls its punches. Pointed at AI, the same design raises the same doubt about who really sets the bar the labs must clear.

For you, the read is plain. Speed and independence rarely come out of the same budget. This design buys speed by letting the industry hold the pen, and how far that tilts toward capture depends on rules that have not been written yet.

Why this is the opposite of Europe's path

The European Union chose public enforcement, not industry self-policing. The AI Act is enforced by public authorities, its obligations on general-purpose models are already biting, and fines can land from 2 August 2026. No industry-funded body sits between a lab and the regulator, the state itself sets and enforces the rule.

So an owner deploying on both sides of the Atlantic faces two divergent regimes on the same model: a US industry review that could speed access, and an EU public track that carries real financial penalties. The two gates will not open at the same moment, and the same model can clear one weeks before the other.

What an owner should watch

Track which model versions clear which gate, and when. A US model may pass an industry review and ship in America before it satisfies the EU's public obligations, widening the availability gap European and UK buyers already manage between markets. That gap is now a planning input, not a surprise.

Nothing here is law yet, and the voluntary first phase means the shape can still change. But the direction, industry self-regulation in the US set against public enforcement in Europe, is the terrain of the next few years. Plan procurement and compliance for divergence, not convergence.

Read next: The AI Rulebook 29 Nations Signed, and the EU Did Not  |  The AI Label Rule Lands on You, Not Your Vendor