Senate negotiators near a bill making AI catastrophic risk a legal duty of care -- while two more sweeping rivals compete for the same floor time

Senate Majority Leader John Thune, Commerce Committee Chair Ted Cruz, and Sen. Amy Klobuchar are circulating draft legislation that would put a legally enforceable duty of care on the handful of companies building the most advanced AI models -- replacing the voluntary safety pledges OpenAI, Anthropic, and Google have operated under until now. "We need to act now and not wait" on what she called commonsense guardrails, Klobuchar told reporters this week. Negotiators could file text as soon as the week of September 14, though as of this writing nothing has actually been made public.

The mechanism described by aides briefed on the draft is narrower than the rhetoric around it. Developers of the largest models would have to design their systems to prevent catastrophic risks -- specifically, AI being used to help someone build a biological or nuclear weapon -- and work with government scientists to test for it before release. If a model fails that test, federal regulators could block it from shipping; the company's only recourse would be a challenge in federal court, not an internal appeal. States that have passed their own AI-safety laws covering that same narrow risk category would see them preempted, the kind of single-national-standard argument the current administration has pushed all year.

That's a different design than the most-cited comparison point, the EU's AI Act, uses. One analysis of the draft reports a specific compute cutoff -- 10^25 floating-point operations, the same figure Brussels uses to decide which models count as general-purpose AI with systemic risk -- but that figure appears in exactly one outlet's account. Every other report of the same negotiations, including wire reporting, describes a bill whose actual coverage threshold isn't public yet, because the bill's text isn't public yet. Worth treating the FLOP number as a plausible guess at what the finished bill might borrow, not a confirmed provision.

The push has momentum for a reason: Cruz's own committee already tried and failed to move AI legislation once this year. A planned July markup of a package of AI bills was postponed so the committee could focus instead on the Kids Online Safety Act, after negotiators couldn't lock down bipartisan support in time. The catastrophic-risk bill now being floated is what came out of the extra two months -- and its chief sponsor is still hedging in public. (Cruz has separately called the risks AI poses "frightening" in television interviews this month, a notably different register from his usual deregulatory framing on tech policy.)

The bipartisan draft is also not the only answer on the table. Two more sweeping proposals are already public, and neither has anything like the same institutional backing.

Sanders and Rep. Greg Casar introduced the more radical of the two, the Ban Artificial Superintelligence Act, on September 3. It would permanently prohibit building or deploying superintelligent AI, temporarily pause advanced development until a new cabinet-level agency sets safety rules, and punish violations with what the senator's own office calls a corporate death penalty for companies and up to 20 years in prison for individuals -- a penalty structure the release explicitly compares to unlawful nuclear-weapons development. "The future of humanity cannot be left in the hands of a handful of Big Tech oligarchs," Sanders said announcing it.

Khanna's plan is less a bill than a proposal he posted publicly rather than filed: a new regulator modeled on nuclear and aviation oversight, mandatory pre-certification, criminal penalties for releasing an uncertified model, mandatory liability insurance for agentic systems, and whistleblower protections for the engineers who build them. Unlike the bipartisan draft or Sanders' bill, it hasn't been introduced as legislative text at all.

Industry's posture toward the three is not uniform, which is itself informative. OpenAI's head of global affairs, Chris Lehane, has said the company wants to work with Congress on a national framework rather than fight one. No frontier model lab has said anything comparable about Sanders' bill, which several members of Congress have already described as having no realistic path in a Republican-controlled chamber this session.

“It is possible, but it ain't easy.” — Sen. Ted Cruz, on the bipartisan bill's chances

The clearest fault line, though, runs inside the bipartisan coalition itself. Sen. Maria Cantwell, the top Democrat on the Commerce Committee and a negotiator on the same bill, has publicly warned that the emerging draft risks becoming "a weak federal standard" that blocks states from enforcing tougher protections of their own -- the exact preemption tradeoff the bill's own backers describe as one of its selling points. That's a Democrat criticizing the bill she's helping write, not an outside objection, and it's the strongest sign yet that the preemption language is genuinely unsettled rather than a done deal dressed up as a negotiation.

Nothing here is fixed until a bill is actually filed, and the calendar isn't generous: the House sits for one week and the Senate for three before members leave town ahead of the November 3 midterms. Cruz's own assessment -- it is possible, but it ain't easy -- reads less like confidence than like a chairman who has already watched one AI markup collapse this year and knows exactly how the second one could too.