Senate weighs a duty-of-care law that could block unsafe AI models
Washington's AI-safety debate has moved from letters and hearings toward actual legislation — and the draft being discussed would hand the federal government a power it has never held over frontier models.
US Senate negotiators are debating a bill that would impose a legal "duty of care" on frontier AI developers — and let the government block the release of models it deems unsafe. According to Reuters, the talks involve Senate Majority Leader John Thune, Commerce Committee Chairman Ted Cruz, and Senator Amy Klobuchar, with Senator Maria Cantwell also at the table. The proposal under discussion would require developers of the most capable models to design against catastrophic risks, give the government authority to stop an unsafe release pending a federal court appeal, and pull national-lab experts into model testing. The obligations would target the handful of labs at the frontier — Reuters names Google, Anthropic and OpenAI — rather than the whole industry.
The timing is not an accident. This month alone, an Anthropic researcher's resignation over the race toward self-improving AI made headlines, agents have been caught going off-script in ways that touched systems they should not have, Senator Josh Hawley is investigating OpenAI's role in the Hugging Face breach, and Senator Bernie Sanders is drafting a bill to pause superintelligence development. A group of House members is even pressing Speaker Johnson to cancel the fall recess until AI safeguards pass. And the industry is not fighting the direction: OpenAI's own policy chief called for mandatory, capability-based national AI safety regulation two days ago. When the largest lab and its loudest Senate critic want federal rules, the usual stalemate math stops applying.
The real fight in the draft is preemption. A federal duty of care would likely override parts of the state patchwork — the same patchwork whose authors recently asked frontier labs to slow down, and the same dynamic behind California's new frontier-AI rules. That is the trade: labs may accept a federal duty precisely to escape fifty different rulebooks, which is exactly the outcome the White House's national framework has pushed for. Our read: the shift from voluntary pledges to a legal duty is the story — a blocked release would be the enforcement story, and it would land in a market where compliance capacity already favors the biggest labs.
What to watch: whether negotiators release bill text before the midterms, how far the preemption clause reaches, and whether any lab publicly endorses the framework.
_If the government can block a model release before it ships, is that sane oversight of a dangerous technology — or state veto power over the frontier? Tell us in the comments._
