Senate AI Duty-of-Care Plan Tests Washington’s Resolve
Bipartisan negotiators are weighing audits and release blocks for frontier models, turning AI safety from industry debate into a possible federal duty.
A possible federal safety obligation
U.S. Senate negotiators are weighing a new framework that would require developers of the most advanced AI models to demonstrate that they are taking reasonable precautions against catastrophic harm. The proposal, reported by Reuters on September 14 and still under negotiation as of September 16, would move beyond voluntary safety commitments by creating a formal “duty of care” for frontier-model companies.
The framework under discussion could give the Commerce secretary authority to demand evidence of safety precautions and deploy government auditors to test companies’ products. Earlier versions of the negotiations also contemplated allowing the federal government to block the release of models judged unsafe, with companies able to challenge those decisions in federal court. The bill’s language, scope and enforcement mechanisms remain unsettled.
The talks involve Senate Majority Leader John Thune, Commerce Committee Chairman Ted Cruz and Democratic Senator Amy Klobuchar, with Senator Maria Cantwell also involved in discussions. That lineup matters: a measure that survives negotiations among senior lawmakers from both parties would have a better chance of clearing the Senate’s 60-vote threshold than earlier, more sweeping AI proposals.
Why it matters
The proposal would change the basic question facing frontier labs. Instead of asking whether a company has published a voluntary safety policy, regulators could ask whether it can prove that its products were tested, monitored and constrained against defined risks before release.
That could affect product timing, model evaluations and the legal exposure of companies including OpenAI, Anthropic, Google and other developers of highly capable systems. Government testing would also create a new institutional role for federal officials, potentially making model release resemble a regulated product launch rather than a purely commercial software decision.
The political timing is significant. The negotiations follow public warnings from current and former AI researchers, calls by several technology leaders for a slower development pace, and a widening debate over whether competition with China justifies limited oversight. Axios reported that proposed House language could give the Department of Homeland Security emergency authority to order companies to slow or shut down systems posing catastrophic risks, suggesting that congressional debate may be moving toward intervention powers rather than disclosure rules alone.
What remains uncertain
No public bill text establishes what “reasonable precautions” or “catastrophic risks” would mean. Those definitions will determine whether the framework targets only extreme biosecurity, cyber or loss-of-control scenarios, or reaches more familiar harms such as fraud, discrimination and unsafe automation.
The administration’s position is another obstacle. President Donald Trump has argued that existing authorities are sufficient and has opposed calls for an AI development moratorium, signaling that even a bipartisan congressional compromise could face resistance. State-law preemption is also unresolved, raising the possibility of a federal standard that limits stricter rules adopted by states.
For now, the development is less a regulatory breakthrough than a test of whether Washington can convert rising anxiety about frontier AI into enforceable obligations before the next generation of models arrives.