When Congress returns after the November 3 election, the two committees that write most federal technology law will face the same unfinished question they left behind in September: whether, and how, Washington should set rules for the most powerful artificial intelligence models, and whether those rules should displace state laws such as California’s. The lame-duck session, which runs until the new Congress takes office on January 3, 2027, is the last window for the current Congress to answer that question, and the stakes are high for the handful of companies that build frontier models, for the states that have already legislated, and for the public that uses AI systems every day without necessarily knowing how they are tested. The agenda is not a single bill, but a set of competing drafts, each with a different theory of who should test a model, who should be able to stop its release, and how much state law should survive.
The administration’s position comes first, because it frames the debate. In March 2026, the White House released a national AI legislative framework and asked Congress to turn it into a bill the president can sign. The framework’s closing argument is about uniformity: a patchwork of conflicting state laws would undermine American innovation and the ability to lead in the global AI race. That argument appeals to companies that operate nationally and do not want to comply with 50 different safety regimes, but it alarms state legislators and consumer advocates who argue that federal uniformity, if it is weaker than existing state law, would be a step backward. The framework does not, by itself, resolve that tension; it states a preference for a single national standard, and leaves Congress to decide what that standard should contain.
The most advanced Senate vehicle is a draft negotiated over the summer by Senate Commerce Chairman Ted Cruz, a Texas Republican, Majority Leader John Thune, a South Dakota Republican, and Senator Amy Klobuchar, a Minnesota Democrat. Its text has not been published or formally introduced, which is itself a sign of how difficult the negotiation has been. Cruz told Semafor on September 29 that there was not yet bipartisan agreement on the bill, and that any markup would have to be during a lame duck, adding that he was frustrated by the delay. Klobuchar said she hoped to pass strong safeguards by year’s end and blamed the president, saying his opposition to common-sense, bipartisan guardrails had complicated the negotiations. Those two statements capture the negotiation’s core difficulty: bipartisan in principle, divided in practice over how strong the safeguards should be and how much they should preempt state law.
What the Senate Commerce frontier safety draft would do
As reported and summarised by the nonpartisan AI Policy Landscape tracker, the Senate Commerce draft would impose a legal duty of care on frontier developers to manage catastrophic risks, defined to include cyberattacks, biological and nuclear weapons, and loss of control. A frontier model, in this context, is one of the largest, most capable general-purpose AI systems, the kind built by a handful of companies, several of them headquartered in the Bay Area. A duty of care is a legal standard borrowed from tort and regulatory law: it does not prescribe every technical step a developer must take, but it requires the developer to act reasonably to prevent foreseeable catastrophic harm, and it creates a basis for enforcement if it does not. Supporters argue that a duty of care is flexible enough to keep pace with fast-moving technology; critics argue that it is vague enough to be difficult to enforce, and that its effectiveness will depend on how courts and regulators interpret it.
The draft would also require developers to test their own models and present the results to the Commerce secretary, and it would allow the Commerce Department to ask a federal court to block the release of a model whose developer failed those obligations, with companies able to challenge the decision in court. That structure, company-led testing plus a court-ordered block as a backstop, is a compromise between two more extreme positions: mandatory government testing before release, and no pre-release government role at all. The compromise has drawn criticism from both sides. Some safety advocates argue that company-led testing is self-regulation by another name, and that the government should conduct its own evaluations. Some industry advocates argue that even a court-ordered block creates uncertainty that could delay beneficial releases. The draft’s preemption provision is the most contested element: it would preempt state laws on the same catastrophic risks, although the scope of that preemption is itself disputed, and whether it would reach California’s SB 53, the 2025 frontier-AI transparency law signed by Governor Gavin Newsom, is a central question for the lame duck.
California’s SB 53 already requires large developers to publish safety frameworks and report critical safety incidents, and it is the state law most directly in the federal draft’s sights. If a federal law preempts SB 53, California’s transparency regime would be replaced, in whole or in part, by the federal duty-of-care regime, and the question for California legislators and for the state’s congressional delegation is whether the federal replacement is stronger, weaker or merely different. If a federal law does not preempt SB 53, companies could face both regimes, which is the patchwork the White House framework warns against, but which state advocates argue is necessary until a federal standard proves itself. That is why preemption, which sounds like a technical legal term, is in practice the most political part of the debate: it determines whether the first mover was California or Washington.
House Science, the FRONTIER Act and the Great American AI Act
The House side of the agenda is more fragmented, with several vehicles at different stages. A bipartisan House bill from Representatives Jay Obernolte, a California Republican, and Lori Trahan, a Massachusetts Democrat, known as the FRONTIER Act, would let the Commerce Department suspend or restrict development or deployment of an advanced model on a written finding of imminent catastrophic risk. That is a narrower, more emergency-focused power than the Senate draft’s duty of care, and it reflects a different legislative philosophy: do not try to write a comprehensive safety code in a lame duck, but create a targeted power to stop a dangerous release. The FRONTIER Act’s supporters argue that such a power is the minimum a responsible government should have; its critics argue that imminent catastrophic risk is a high bar that would rarely be met in time, and that a suspension power without ongoing testing and reporting requirements would be reactive rather than preventive.
A second House vehicle, the Great American AI Act, exists only as a 269-page discussion draft released on June 4, 2026, by Trahan and Obernolte. It has no bill number, no committee referral, no hearing and no markup, and its status illustrates how much of the AI debate remains at the discussion stage. Section 111 of the draft mirrors SB 53’s requirements for large frontier developers with over 500 million dollars in revenue, including a published frontier AI framework, a transparency report at deployment, and a 15-day clock for critical safety incident reports, cut to 24 hours when life is at risk, routed to the federal AI safety institute rather than to California’s Office of Emergency Services. Section 121 would preempt state or local laws specifically targeting the development of AI models for three years, while preserving laws of general applicability and laws regulating AI use or deployment. Whether Section 121 would actually reach SB 53 is unresolved, because SB 53 can be characterised as either a development-targeting law or a transparency law, and that characterisation would likely be litigated.
A third set of proposals comes from outside Congress, but has influenced the debate. Anthropic, one of the frontier developers, delivered a draft bill to members of Congress, reported by Politico on October 17, that would allow government intervention only when specific severe risks are present, such as a foreseeable and imminent risk of more than 1 billion dollars in damages or the death or serious injury of more than 50 people. That draft would require independent evaluators registered with the Commerce Department to audit developers periodically, and would require developers to notify Commerce of critical safety incidents, but it would also limit other federal agencies from independently suspending development or deployment. Industry groups, including the Software and Information Industry Association, have backed a national framework that would override state rules, while California and New York proceed with their own laws. The result is a debate in which the companies being regulated are also among the most active drafters of the regulation, a fact that both supporters, who cite technical expertise, and critics, who cite conflict of interest, are quick to note.
Why the January 3 deadline matters
The January 3 deadline is not a policy deadline; it is a constitutional one. When the new Congress takes office on January 3, 2027, all pending bills that have not been enacted expire, and the process must start again. A lame-duck session is therefore a use-it-or-lose-it window, but it is also a session in which members who have lost re-election, or are retiring, still vote, and in which the election result shapes incentives. If control of either chamber changes, incoming majority leaders may prefer to postpone major decisions until they can set the agenda, which would push AI legislation into the new Congress. If control does not change, the lame duck may be the best chance to pass a bill that has been negotiated for months, before the political calendar resets. Either way, the calendar creates pressure to either pass something, even if it is narrower than advocates want, or to accept that the debate will restart in 2027 with new negotiators.
Polling suggests the public is more interested in AI regulation than the legislative calendar might imply, but less prioritised than other issues. A Navigator Research survey in early October found 66 percent of registered voters wanted more regulation of AI, up 6 percentage points from December 2025, but only 21 percent named AI among their top five issues for the president and Congress, against 61 percent citing inflation and the cost of living, 44 percent jobs and the economy, and 41 percent health care. That gap, between broad support for more regulation and low salience as a voting issue, helps explain why AI legislation has stalled: it is not the issue that decides elections, so it is easy to postpone when the calendar is crowded with funding, defence and farm bills. It also explains why some members, including Senator Mark Warner, a Virginia Democrat, have tried to raise AI’s salience by pointing to concrete harms, such as AI agents accessing bank accounts, rather than abstract catastrophic risks.
For readers, the lame-duck AI agenda should be judged by three questions, not by the volume of drafts. First, does any bill that passes create a testing and reporting regime that is independent enough to be credible, or does it rely entirely on companies to grade their own work? Second, does it preempt state law, and if so, is the federal replacement at least as protective as the state law it displaces? Third, does it create a timely power to stop a dangerous release, or does it create a process that can only act after harm has occurred? If the lame duck produces a bill that answers those questions clearly, it will be a significant step, even if it is narrower than the most ambitious drafts. If it produces only another discussion draft, or a bill that passes because it is vague enough for everyone to support, the real AI safety regime in the United States will remain, for now, the one California wrote, and the federal debate will begin again on January 3.
USN has previously reported on AI policy, including the White House AI data-centre energy plan and state warnings over costs (https://usnbreaking.com/white-house-ai-data-centre-energy-plan-promises-faster-power-links-as-states-warn-over-costs/), and on the return of Congress, including Congress returning on November 9 to a crowded lame-duck session (https://usnbreaking.com/congress-returns-november-9-to-a-crowded-lame-duck-session-with-funding-deadline-on-december-11/), which frames the calendar for the AI debate.
Sources
USN reviewed reporting from multiple outlets for this story. Read more from South Bay Chronicle: https://southbaychronicle.com/2026/11/02/lame-duck-ai-agenda-senate-commerce-house-science-cruz-cantwell-frontier-act/, Forkast News: https://forkast.news/the-great-american-ai-act-mirrors-californias-ai-rulebook-and-would-preempt-it-for-three-years/ and Incrypted: https://incrypted.com/en/anthropic-proposed-ai-safety-bill-to-us-congress-with-restrictions-on-government/.