Senate AI Bill Hits Dispute Over Who Tests Powerful Models

A bipartisan Senate effort to regulate the most powerful AI models has run into a basic question of trust: should developers test their own systems, or should federal experts do it before the models are deployed? Senate Commerce Committee negotiators are divided over that choice. The bill remains unreleased, and its path is uncertain.

The proposal is being developed by Sens. Amy Klobuchar and John Thune, with input from committee chair Ted Cruz. A Democratic committee aide told Nextgov/FCW that draft provisions would have companies conduct model safety tests and submit the results to the Commerce secretary for approval to deploy a model. The bill text is not public, so the reported framework remains subject to negotiation.

The disagreement is over who validates the test

Maria Cantwell, the committee’s ranking Democrat, wants mandatory testing and security vetting by federal entities, including national laboratories and national-security agencies. She has argued that scientists and specialists should assess whether frontier models could enable sophisticated cyberattacks or assist biological or nuclear weapons development.

That difference is more than a procedural detail. The company-run approach would make developers responsible for the initial evaluations, with the Commerce Department reviewing their submissions. Cantwell’s approach would put government scientists and security experts closer to the testing itself. Her office has previously questioned whether developers are best positioned to assess AI risks.

Safety powers and state rules remain unsettled

• The bill has been described as requiring safety testing and incident reporting for the most powerful AI models.
• It could give the Commerce Department and Department of Homeland Security authority to block a model’s release if it is judged to pose a genuine catastrophic risk, including biosecurity or nuclear risks.
• The federal framework could preempt state AI laws, according to Fortune’s account of the proposal.

The possible effect on state law is another source of friction. Cantwell has said the proposal could create a weak federal standard that undermines existing state legislation. Cruz, meanwhile, has said the coming bill will address catastrophic frontier-AI risks while preserving the United States’ ability to lead China in AI development and deployment.

An unreleased bill, with industry in the room

The negotiations have not produced a public legislative text. A markup planned before the August recess was canceled, and Fortune reported that OpenAI and Anthropic were privately weighing in on drafts with Senate staff. Nextgov/FCW also reported that, according to a person familiar with the talks, Anthropic and AI safety groups opposed or would not support the bill’s current approach; Anthropic did not comment to the outlet.

For now, Congress is debating the mechanics of a system it has not shown the public. The eventual answer will determine whether federal oversight means checking a developer’s safety file, conducting an independent evaluation, or some combination of both.