Behind closed doors, three senior senators are negotiating a measure that would establish new guardrails for advanced artificial intelligence systems in the United States. The proposed legislation would impose a duty of care on AI developers and grant Washington authority to halt the release of models deemed to pose unacceptable risks. According to reporting by Courtney Rozen for Reuters, the draft would also preempt state-level AI regulations. The central point of contention: which institution should conduct the critical safety evaluations.
Duty of care and government veto power
The emerging bill would require AI developers to implement safeguards designed to mitigate catastrophic hazards, according to two Senate aides involved in the process. The legislation would simultaneously authorize federal officials to prevent the deployment of models that fail safety assessments. Companies dissatisfied with such determinations could pursue legal remedies in federal court, though the extent of governmental authority remains under discussion. The measure would apply specifically to the most capable models—a category that encompasses systems developed by Google, Anthropic and OpenAI.
Who is driving the negotiations
Three senators are steering the talks: John Thune, the Senate majority leader; Ted Cruz, who chairs the Commerce Committee responsible for overseeing the Commerce Department and its AI safety research division; and Amy Klobuchar, the ranking Democrat at the table.
Cruz announced on X that he is collaborating with Klobuchar and Thune on addressing catastrophic risks tied to biological or nuclear weapons. Earlier this month, he indicated the committee might proceed to a markup session, having previously postponed one scheduled before the August recess to prioritize child safety measures.
Klobuchar stated in a prepared remark that any final agreement must obligate developers to partner with government specialists to validate and evaluate models. She emphasized that the nation cannot tolerate the release of dangerous systems, particularly those that might escape their creators' control. Thune's office declined to provide comment to Reuters.
Testing authority emerges as the flashpoint
Maria Cantwell, the senior Democrat on the Commerce Committee, is not part of the negotiating trio—and her absence signals deeper disagreement. Alexandra Kelley reported for Nextgov that Cantwell is withholding support over the proposed testing structure.
Under the current draft, as described by a Democratic committee aide, companies would conduct their own safety assessments and submit results to the Commerce secretary for deployment authorization. Cantwell envisions a different path: models would be forwarded to federal agencies, including the national laboratories and national security entities. The aide characterized the existing proposal as "primarily a voluntary standard type situation". Cantwell's office has submitted alternative language to Klobuchar, whom the aide described as open to revision. The decisive factor will be Cruz's position.
A second source familiar with the deliberations told Nextgov that Cantwell, Anthropic and AI safety organizations are "refusing to play ball". Anthropic did not respond to Nextgov's inquiry.
Cantwell articulated her stance on X, stating that substantive legislation should route the most powerful models to scientists at the national laboratories for assessment of whether they could facilitate advanced cyberattacks or accelerate biological or nuclear weapons development.
A provision that would override state authority
The bill includes language that would prevent states from enforcing their own regulations concerning certain model-related hazards, according to the two Senate aides. One aide cited a scenario: a malicious actor leveraging an AI system to engineer nuclear or biological weapons.
This preemption matters significantly because states have been the primary regulatory arena. California enacted SB 53, which OpenAI initially opposed before requesting the state strengthen it. New York passed the RAISE Act. Colorado's legislature dismantled its own law this year following litigation.
Cantwell linked the preemption clause to earlier legislative battles. She wrote on X that Republicans are now recognizing the necessity for stringent AI safety rules, and she welcomes this shift in urgency. She characterized it as a dramatic departure from the previous push for a decade-long moratorium on state AI regulation. The Senate voted 99 to 1 in July 2025 to strip that moratorium from legislation.
Federal protocols for evaluating frontier models already exist, though they remain confidential. The Next Web reported on 9 September that the White House has not made its framework public, and that only the largest companies have been granted access to review it.
Time pressure from the election calendar
The midterm elections fall on 3 November. The Senate is anticipated to remain in Washington for approximately three additional weeks. The House is scheduled for one week, with members returning on 9 November.
Four House Democrats have petitioned the Speaker to alter this schedule. Sam Liccardo drafted a letter signed by George Whitesides, Lori Trahan and Ted Lieu. The correspondence urges the House to reconvene immediately and maintain continuous sessions until bipartisan AI safeguards advance. Andrew Solender obtained the letter for Axios.
The letter references bills already introduced, including the AI Kill Switch Act, co-authored by Lieu and Nathaniel Moran, and the FRONTIER Act, authored by Trahan and Jay Obernolte.
Mike Johnson responded on NBC's Meet the Press on 13 September, stating he would bring the House back and schedule a vote if a legislative solution existed to vote on. He is advocating for a summit involving the president, congressional leadership and representatives from the major AI laboratories. Instead, Axios reported, he is advancing a data centre energy bill this week.
The White House is not pushing for action
When asked in Dallas last Thursday about concerns regarding AI's existential dangers, Donald Trump stated he harbored none. Speaking in Ireland on Sunday, he dismissed the warnings as originating from negative actors promoting scenarios that will not materialize, and asserted that dominance in AI determines broader success. The Next Web covered these remarks on Monday.
David Sacks, Trump's former AI czar, has instructed the labs to exercise restraint without legislative intervention. The Next Web reported this position on Sunday. Johnson echoed the same argument during his Meet the Press appearance.
What remains uncertain
The bill's text has not been shared publicly. This leaves three critical questions unanswered: the breadth of the government's authority to block models, the mechanism for judicial review of blocking decisions, and the geographic scope of the preemption clause.
Klobuchar and Thune co-authored legislation in 2023 titled the Artificial Intelligence Research, Innovation, and Accountability Act, which failed to advance. A Democratic aide told Nextgov that the current draft draws inspiration from that earlier effort but represents a fundamentally distinct bill.
Source: The Next Web
