Senate Commerce Nears Markup of AI Safety Legislation as State Laws Advance

Senate Commerce Committee chair Ted Cruz says his panel could mark up AI safety legislation before the chamber's calendar tightens this fall, a step that would move the first binding federal rules for frontier model developers closer to a floor vote. The Texas Republican has endorsed legislation aimed at preventing catastrophic risk and has described recent safety warnings from departing AI researchers as highly concerning.

Cruz is negotiating the package with Majority Leader John Thune and Senator Amy Klobuchar. Draft text is circulating among AI labs and public advocacy groups that are already sending feedback to committee staff. California added pressure to the federal talks this week by enacting the first US state law governing how independent auditors evaluate AI products.

What the Senate Is Negotiating

The Cruz, Klobuchar and Thune bill has not been released publicly. As drafted, it would give the Commerce Department and the Department of Homeland Security leverage over frontier models while overriding state AI safety statutes. Thune and Klobuchar have worked from a duty of care model built on developer self-testing, in which labs document and test their own systems against defined risk thresholds before release.

Cantwell, the committee's lead Democrat, is pushing a stricter version. Her approach would place enforceable legal duties on AI companies to manage catastrophic risk, and she has resisted provisions that weaken state protections. The gap between self-testing and enforceable duties is the central disagreement in the negotiation.

The distinction is whether the federal statute sets a floor that states may build on or a ceiling that caps them. The two models also carry different costs. Self-testing keeps release cycles fast and places the compliance burden inside the lab, which is cheaper for developers and easier for a committee to pass. Enforceable duties open companies to liability and require an evidentiary record, which slows deployment and raises the price of a bad release. California's auditor statute leans toward the second model by creating a role for outside evaluators rather than accepting a developer's own findings.

Scope is the other unresolved question. A bill keyed to biological and nuclear threats captures a narrower set of models than one covering general catastrophic risk, and the capability thresholds that trigger obligations determine which labs face testing at all. Set the bar high and most deployed systems stay outside the regime. Set it low and mid-sized developers absorb compliance costs that the largest labs can spread across far more revenue.

The security framing explains why the Department of Homeland Security appears in several drafts alongside civilian agencies. Biological and nuclear scenarios are the stated trigger for the strictest obligations, which pulls the text toward a security-agency model with classified threat assessments rather than a consumer-protection model. Security agencies can restrict disclosure of testing details, which limits what outside auditors and independent researchers can verify.

How the Competing Proposals Compare

AI Safety Legislation Hits the Preemption Wall

Preemption draws the sharpest opposition. A federal framework that displaces state AI safety laws would undercut California's auditor statute in the week it was enacted and freeze similar bills advancing in other states. The Center for Democracy and Technology has argued that a package centered on children's privacy and AI cannot justify overriding state AI rules, because the federal text covers a narrower set of harms than the state laws it would replace.

NetChoice, whose members include major technology platforms, has urged the committee to reject top-down mandates, contending they would create legal traps and sweeping censorship. It wants Congress to pursue digital literacy, parental controls and free enterprise instead. Civil rights organizations make the opposite complaint: the Leadership Conference on Civil and Human Rights called the committee's August markup a missed opportunity, because the bills extend privacy protections to users under 18 but leave data practices affecting adults unaddressed. Its polling found 81% of respondents who expect AI to help them still want companies required to prove their systems do not discriminate.

For the largest labs, a single federal standard is the cheaper outcome: one compliance program instead of fifty, and a preemption clause removes the risk of conflicting obligations across jurisdictions. For smaller developers, and for states that have already invested in oversight, the same clause removes protections without guaranteeing an equivalent federal replacement, since the Senate text has not been released and its enforcement mechanism is not public.

The Other Vehicles in Play

Separate bills are moving on their own tracks. Representatives Ted Lieu and Nathaniel Moran have introduced a bipartisan measure requiring frontier developers to keep the technical capacity to throttle, suspend or shut down their most advanced systems, and it would authorize the Homeland Security Secretary to order a slowdown or shutdown when a model could cause catastrophic harm. Senator Mark Warner's Secure AI Development Act would impose binding pre-deployment testing on frontier developers, one of the first statutory testing requirements at the federal level if it becomes law.

In the House, the FRONTIER Act would cover the top of the risk pyramid, with companion bills handling downstream harms such as automated denials of benefits and civil rights violations. House and Senate leaders have not reconciled those frameworks, and two House committees hold overlapping jurisdiction over the frontier text, which adds a scheduling obstacle independent of the Senate's own calendar.

Senate Commerce has already cleared part of its stack. The committee approved the Kids Online Safety Act and several children's AI privacy measures at an Aug. 5 markup, after an earlier July 29 target slipped. The frontier safety provisions were not part of that vote.

States Move While Washington Negotiates

State-level organizing has accelerated. A new policy platform launched this week will track emerging AI bills nationwide and supply policy expertise to state lawmakers, treating statehouses as the primary venue for AI safety legislation while Congress negotiates. For developers, that creates a dual compliance track: a federal framework that may preempt state law, and a growing set of state statutes that apply until it does.

The practical burden lands on compliance teams at the labs and on the enterprises that buy from them. A developer that already runs internal red-team testing can map those results to a self-testing standard with modest changes. Third-party audit rules require new documentation, outside evaluators and review timelines that push back release dates. Enterprise buyers inheriting those artifacts will need to know which regime governs a given deployment, particularly if a state statute and a federal standard diverge during a transition period.

The political calendar narrows the window further. With midterm elections approaching, lawmakers in both parties have acknowledged that existing law was not written for systems that act autonomously, and proposals that stall this fall are unlikely to return before the next Congress. Committee schedules have already slipped once, from a July 29 target to August.

Why this matters

The markup date for AI safety legislation matters less than the preemption clause. Congress has spent more than a year circulating frontier AI drafts without a floor vote, while California enacted an auditor law and other states drafted their own. If Commerce advances a package that overrides those statutes, developers get one federal standard and lose the state-by-state flexibility they now work under. If preemption is stripped to win Cantwell's support, the federal bill adds obligations on top of a state patchwork instead of replacing it. The signals to watch are whether Cruz schedules the markup before the recess and whether preemption language survives the committee print.

Sources

Warner Rolls Out Comprehensive AI Legislative Agenda Focused on Responsible Innovation, Workers, and National Security

Related Articles

• California AI Audit Laws Signed as Newsom Pushes for National Rules
• Bipartisan Proposal for Great American AI Act Establishes National Oversight Standards
• U.S. Commerce Department Secures AI Safety Agreements with Google, Microsoft, and xAI

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.