There were eleven people in the gallery for Tuesday morning's Senate Commerce subcommittee hearing on "foundation-model evaluation standards." Four of them were Hill staffers. Two were trade-press reporters. The remaining five were lobbyists, and they were the ones taking notes.

What was actually being argued

Beneath the procedural surface, the hearing was a proxy fight over a question Congress has spent two years avoiding: who decides what counts as a "frontier" model, and what compliance burden attaches once a system clears that line.

The administration's preferred answer involves a federal evaluation institute housed at NIST, voluntary commitments backed by procurement leverage, and disclosure rules pegged to training-compute thresholds. The industry's preferred answer involves the same institute, weaker disclosure rules, and an explicit federal preemption of state-level AI legislation. The civil-society groups want the disclosure rules and the institute but oppose preemption.

Preemption is the whole ballgame

California's SB-1047 successor, currently in conference committee, and the New York EQUITY Act are the two state laws the industry most wants blocked. Both impose obligations that go beyond anything in the pending federal drafts — third-party audits for models above defined capability thresholds, whistleblower protections for lab employees, and, in New York's case, a private right of action for downstream harms.

A federal law that preempts state action while imposing lighter federal requirements would be, from the industry's standpoint, a win. A federal law without preemption would be, from the industry's standpoint, an additional layer of compliance stacked atop the state ones.

What the ranking member said, and what she meant

Senator Kaine's questions, delivered with the polite indirection of someone who knows the transcript matters more than the room, walked through the compute-threshold definitions line by line. Her staff, per two people involved in the negotiations, has been drafting an alternative that would peg the definition to capability benchmarks rather than compute inputs.

"If we tie the rules to floating-point operations, we're regulating last year's architecture. In three years the frontier will fit under whatever number we pick today."

— Senator Tim Kaine, in the subcommittee's second round of questions

What to watch

The subcommittee markup is scheduled for the third week of July. The full committee will not move until the September work period at the earliest. The White House has, so far, not signaled whether it will insist on the compute-threshold formulation. If it doesn't, the fight over what replaces it will consume the fall.

About the author

David Okafor

Politics Editor in Washington. Two decades on the Hill.