WASHINGTON — When Congress returns after Tuesday’s midterm elections, one of the most consequential technology questions left unfinished this year will move to the front of a very crowded lame-duck calendar: whether the federal government should set binding safety rules for the most powerful artificial intelligence models, and whether those rules should override state laws that already try to do the same job. The debate is not about chatbots in general, or about the everyday tools millions of people now use at work and school. It is about a much smaller group of systems — the largest, most capable general-purpose models built by a handful of companies — and about the catastrophic risks that critics say those systems could create if they are misused, fail, or escape meaningful control.
For months, that question has been negotiated mostly behind closed doors in the Senate Commerce, Science and Transportation Committee, the panel that writes much of federal technology law. Its chairman, Senator Ted Cruz of Texas, Senate Majority Leader John Thune of South Dakota and Senator Amy Klobuchar of Minnesota spent the summer working on a frontier-safety draft. By late September, however, the talks had stalled. Cruz told Semafor at that point that there was not yet bipartisan agreement on the bill, and that any committee markup would therefore have to wait for a lame-duck session. He said he was frustrated by that delay. Klobuchar, for her part, said she hoped strong safeguards could still pass by the end of the year, and she blamed presidential opposition to what she called common-sense, bipartisan guardrails for complicating the negotiations.
That framing matters because it captures the two pressures now shaping the lame duck. On one side is a sense, shared by lawmakers in both parties, that the absence of a federal standard cannot continue indefinitely while states move ahead on their own. On the other is a White House that has favoured voluntary industry commitments over binding federal mandates, and a calendar that leaves very little room for a complex bill to be written, debated, amended and passed in both chambers before the current Congress ends. The lame-duck session runs from mid-November until the new Congress is sworn in on January 3, and it already has to handle government funding and the annual defence authorisation bill. Any measure that does not pass by January 3 dies and must be reintroduced from scratch.
Congress AI Bill Lame Duck: What the Senate Commerce Frontier Safety Draft Would Do
The Senate Commerce draft has not been published or formally introduced, so any description of it rests on reporting and summaries rather than on statutory text that the public can read line by line. That caveat is important, and it is one reason the debate has been so difficult to settle: lawmakers, companies and advocacy groups are arguing in part about a proposal whose precise language has not been made public. According to summaries of the draft, however, several core ideas have been consistent across the reporting. The first is a legal duty of care for frontier developers — an obligation to manage catastrophic risks, defined to include cyberattacks, biological and nuclear weapons risks, and loss of control over a model.
The second idea concerns testing. Under the draft as described, developers would test their own models and present the results to the secretary of commerce. That self-testing model is central to the disagreement that has slowed the talks. Supporters of the approach argue that the companies building the models understand them best, and that a government-run testing regime would be slow, technically difficult and vulnerable to political pressure. Critics argue that self-testing, without independent verification, is closer to a voluntary standard dressed up as a requirement, and that the public should not have to trust the same companies that profit from a model to certify that it is safe.
The third idea is an enforcement mechanism with a courtroom check. The Commerce Department would be able to ask a federal court to block the release of a model whose developer had failed to meet its obligations, and the company would be able to challenge that decision in court. In practical terms, that would give the government a way to intervene before a model reaches the public, rather than only punishing harm after it has occurred, while preserving a judicial route for a company that believes the government has acted wrongly. How often such a power would actually be used, and how quickly a court could decide a case involving a fast-moving technology, are questions that the draft’s unpublished details would have to answer.
The fourth idea, and in many ways the most contested, is preemption. The draft as reported would preempt state laws covering the same catastrophic risks. Preemption means that a federal law overrides state laws in the same field. In this case, the provision would likely reach parts of California’s SB 53, the 2025 frontier-AI transparency law signed by Governor Gavin Newsom, which requires large developers to publish safety frameworks and report critical incidents to the state. It could also affect other state measures, including New York’s RAISE Act, to the extent that they regulate the same model risks. Supporters of preemption say a patchwork of conflicting state rules would undermine American innovation and the country’s ability to lead in the global AI race. Opponents say preemption would strip states of protections they have already enacted, and replace them with a federal standard that may be weaker, narrower, or slower to enforce.
The Testing Fight: Self-Testing Versus Independent Audits
If preemption is the most politically explosive question, testing is the most technically consequential. Senator Maria Cantwell of Washington, the Commerce Committee’s ranking Democrat, released her own six-principle framework on October 7, and its central difference from the Cruz draft is precisely who does the testing. Under Cantwell’s version, the National Institute of Standards and Technology would write clear, transparent and enforceable safety standards, and covered models would not be released until they had undergone an independent audit confirming that they met those safeguards and standards. Cantwell has compared those audits to outside accountants examining a public company’s financial statements — a check that is credible precisely because it is not performed by the company itself.
Cantwell’s framework also calls for incident reporting, whistleblower protections and a United States-China channel for AI emergencies, sometimes described as a red phone for AI crises. Notably, it does not address preemption, which leaves open the possibility that a federal standard could sit alongside state laws rather than replacing them. That silence is itself a statement: for Cantwell and her allies, the priority is enforceable federal safety standards, and the question of whether those standards should displace state law can be negotiated separately, or left for a later fight. For negotiators in the Cruz-Thune-Klobuchar talks, by contrast, preemption has been presented as part of the bargain — a uniform national standard in exchange for binding federal rules.
The testing dispute is not merely academic. Frontier models are enormously complex systems, and evaluating them for catastrophic risk involves probing how they might behave when deliberately misused, how they might be combined with other tools, and how they might fail in ways their creators did not anticipate. Independent evaluators argue that they bring fresh eyes, different incentives and specialised expertise. Company-led evaluators argue that they have the deepest access to the model, its training data and its internal workings, and that outside auditors may lack the tools or the time to do a thorough job before release. A workable law would have to decide not only who tests, but what counts as a pass, what happens when a test is inconclusive, and how much of the testing must be disclosed to the public, to regulators, or to no one at all.
That last point — disclosure — connects the Senate debate to a broader public concern. Many of the risks at issue are not visible to ordinary users. A person chatting with an AI assistant cannot tell, from the conversation alone, whether the underlying model has been tested for its ability to help design a biological weapon, or whether it has safeguards against being manipulated into assisting a cyberattack. Transparency requirements, such as the model cards and risk-management frameworks proposed in the House bill, are designed to close that gap. But transparency also raises trade secrets, security and competitive concerns, and lawmakers have struggled to draw a line that gives regulators and the public enough information without handing sensitive technical details to bad actors.
House Science and the FRONTIER Act: A Parallel Path
While the Senate has negotiated, the House has moved on a parallel track. Representatives Jay Obernolte of California, who chairs the Science Committee’s research and technology subcommittee, and Lori Trahan of Massachusetts introduced the FRONTIER Act on July 23. The bill grew out of a broader discussion draft released in June, and it sets tiered requirements based on developer size: model cards, which are standardised disclosures about how a model was built and tested; risk-management frameworks; independent audits; and incident reporting. Its sponsors describe it as creating a uniform national standard for transparency, auditing and reporting of catastrophic risk, precisely to prevent a patchwork of state regulations.
The FRONTIER Act’s tiered design is an attempt to answer a fairness concern that runs through the whole debate. Not every company building AI is a frontier lab with billions of dollars in computing resources. Smaller developers and open-source projects argue that rules written for the largest models could crush them if applied without regard to scale. By calibrating obligations to developer size, the House bill tries to focus the heaviest requirements where the risks are thought to be greatest, while leaving lighter-touch rules for smaller players. Critics, including some Democrats, have argued that the bill’s preemption provisions make it a nonstarter, because they would override state protections. Supporters, including some Democrats who have praised it as an important bipartisan step, argue that a flawed but real federal standard is better than continued fragmentation.
The House Science, Space and Technology Committee itself is chaired by Representative Brian Babin of Texas. Its role in the lame duck will depend in part on whether House leadership sees the FRONTIER Act as a vehicle that can pass, or as a negotiating position for the next Congress. Committee leaders have in the past been reluctant to rush complex legislation, and the bill’s future may hinge less on its technical merits than on whether party leaders in both chambers decide that an AI bill is a priority for the final weeks of the session, or a topic to carry into the new year with new committee chairs and a new political balance.
The White House Framework and the Voluntary Commitment
Overlaying all of these congressional efforts is the administration’s own position. In March 2026, the White House released a national AI legislative framework and asked Congress to turn it into a bill the president could sign. The framework’s closing argument was about uniformity: a patchwork of conflicting state laws, it said, would undermine American innovation and the country’s ability to lead in the global AI race. It also asked for parental controls for minors, protections against data-centre costs falling on ratepayers, faster permitting for on-site power, and guardrails against AI being used to censor lawful speech. It was, in essence, a legislative wish list rather than proposed statutory text, and it left many trade-offs unresolved.
On September 29, President Donald Trump hosted AI executives at the White House. Asked why they should be trusted to regulate themselves, he answered, according to Roll Call, that they were outstanding people. The companies signed a Joint Commitment on Frontier Responsibilities, which said that over time it may make sense to codify those steps into laws or regulations. For supporters of voluntary approaches, that commitment is evidence that the industry takes safety seriously and that government should build on cooperation rather than impose rigid mandates. For critics, including many congressional Democrats, it is evidence of the opposite: that without binding law, safety commitments remain promises that can be revised, delayed or abandoned when commercial pressures change.
That disagreement helps explain why the September 29 floor episode mattered so much. On the same day as the White House event, Senator Mark Warner of Virginia asked for unanimous consent to pass a bill with Senators Brian Schatz of Hawaii and Andy Kim of New Jersey. The bill would have created an AI Safety Board within the Commerce Department, required frontier developers to give the board access to new models at least 45 days before release, and required safety incidents to be reported within 30 days, or within 72 hours if an incident posed an imminent national-security threat. Cruz objected, saying the bill would put too much authority in the hands of the board and the Commerce Department, and that he would prefer a role for independent judges. He also argued that Congress must not legislate hastily or in a closed manner, and that federal standards must address the growing patchwork of state laws. Schatz countered by asking whether companies heading toward the largest initial public offerings in human history could be trusted to regulate themselves voluntarily.
The objection did not kill the underlying ideas, but it did close off the fastest procedural route. Unanimous consent allows a bill to pass quickly if no senator objects; a single objection stops it. The sponsors can still pursue the bill through regular order — committee hearings, markup, floor debate and votes — but that path takes time, and time is precisely what the lame duck lacks. It also signalled, publicly, the depth of the divide over how much authority any new AI body should have, and whether that authority should sit with an executive agency, a court, or some combination of the two.
Why Passage This Year Looks Unlikely, and What Comes Next
Analysts who track the issue closely have been sceptical about the chances of a comprehensive bill passing before January. The DC/ai Decoded newsletter concluded in early October that AI legislation was unlikely to pass this year. That assessment reflects more than the policy disagreements. It reflects the calendar, the competing demands of funding and defence legislation, the need to reconcile any Senate product with the House FRONTIER Act, and the political uncertainty created by the midterms themselves. Tuesday’s results will decide who chairs the Commerce and Science committees when the new Congress begins, and that, in turn, will shape which framework — the Cruz draft, Cantwell’s version, the FRONTIER Act, or something new — becomes the starting point in 2027.
None of that means the lame duck will be silent on AI. Lawmakers may still use the session to hold hearings, publish draft language, build coalitions, or attach narrower provisions to larger bills that must pass. They may also use oversight tools, such as the November 6 deadline set by Representative Zoe Lofgren of California, the Science Committee’s ranking Democrat, for the Commerce and Energy departments to answer oversight requests, and her joint request with other members that frontier labs report safety incidents to Congress within 48 hours. Such steps do not create law, but they shape the record on which the next Congress will act, and they keep pressure on both companies and the administration to explain their safety practices in public.
For the public, the most important distinction to keep in mind is between a draft, a framework, a commitment and a law. A draft that has not been published is not a law. A framework that sets out principles is not a law. A voluntary commitment signed at a White House event is not a law. Only a bill that passes both chambers and is signed by the president — or passes over a veto — becomes law. Until that happens, the rules that actually govern frontier models in the United States will remain a mixture of state laws like California’s SB 53, sector-specific regulations, contractual obligations, and the companies’ own policies. That mixture may be enough for some, and dangerously incomplete for others, but it is the reality that the lame duck inherits.
The stakes of getting the next step right are high on both sides of the argument. Supporters of strong federal rules warn that a catastrophic misuse or failure, if it occurs before rules are in place, will be far harder to address after the fact, and that public trust in AI could collapse. Supporters of a lighter-touch approach warn that overly rigid rules could slow American innovation, push development overseas, and hand an advantage to competitors, including China, that do not face the same constraints. Both warnings may contain some truth, which is why the negotiation has been so difficult: it is not simply a choice between safety and innovation, but a search for a design that can deliver both, with testing that is credible, enforcement that is fair, and a federal-state balance that neither leaves states defenceless nor fragments the national market beyond repair.
As Congress returns, the question for voters and readers is therefore not only whether a bill passes, but what kind of bill, if any, emerges from the lame duck’s compressed timetable. A narrow transparency bill might pass where a comprehensive safety bill cannot. A bill that sets standards but leaves preemption for later might find broader support than one that tries to settle everything at once. Or the session may end with no AI bill at all, leaving the debate to a new Congress with new leaders, new majorities and the same unresolved questions. What is certain is that the issue will not disappear: the models will keep getting more capable, the states will keep legislating, and the pressure for a federal answer will keep growing, whether the lame duck supplies it or not.
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/), CIO Dive (https://www.ciodive.com/news/ai-regulation-push-gains-steam-ahead-midterms/872929/) and DLA Piper (https://knowledge.dlapiper.com/dlapiperknowledge/globalemploymentlatestdevelopments/2026/US-federal-white-house-releases-the-national-policy-framework-for-artificial-intelligence).