US AI Regulation Hits a Wall: Why the Bipartisan Thune Bill Stalled — and What Happens Next

The Short Answer

The Thune AI bill's stall is a setback for US AI governance — but not the end of the road. Here's what it means:

Federal AI safety legislation is delayed, not dead. The bill will return after the August recess, and the bipartisan interest in AI safety regulation remains real. But the specific provisions that caused the stall — Anthropic's push for public disclosure, Cantwell's concerns about the framework — will need to be resolved, and that resolution may take months, not weeks.

The US still has no federal AI safety law. Unlike the EU (which has the AI Act), China (which has comprehensive AI regulations), and several US states (which have passed their own AI laws), the federal government has not enacted AI safety legislation. The Thune bill was the closest the US has come. Its stall means the regulatory patchwork continues.

AI companies are operating in a regulatory vacuum — and some prefer it that way. The disagreements that stalled the bill reveal a split: some AI companies (Anthropic) want stronger transparency requirements, while others are comfortable with the status quo. The absence of federal regulation benefits companies that prefer self-governance over government oversight. The presence of 1,300 employees calling for government pacing tools suggests self-governance isn't working.

Your practical takeaway: Don't wait for US federal AI regulation to set your AI safety standards. The EU AI Act is already in force. State-level AI laws are proliferating. Major enterprise customers are imposing their own AI safety requirements on vendors. The smart approach is to build AI governance practices that meet the highest current standard (likely the EU AI Act) and adapt as US federal requirements eventually arrive.

What the Thune Bill Would Have Done

The proposed legislation, formally titled the 'Frontier AI Safety and Security Act,' contained several significant provisions:

Legal duty to manage catastrophic risks: AI companies developing the most powerful models would have been legally required to identify, assess, and preemptively manage risks that could cause catastrophic harm — specifically including cyberattacks on critical infrastructure, development of biological or chemical weapons, and loss of human control over AI systems. This is a meaningful legal obligation, not a voluntary guideline.

Mandatory capability reporting: AI companies would have been required to notify the Commerce Department when their models crossed specific capability thresholds — autonomous replication, self-improvement, and other dangerous capabilities defined in the bill. This notification requirement was designed to give the government visibility into what frontier AI labs are building, before those capabilities are deployed.

Court injunction power: The Commerce Department would have been empowered to seek federal court injunctions blocking the deployment of AI models deemed to pose an unacceptable risk. This is the most consequential enforcement mechanism in the bill — the government could go to a judge and say 'this model is too dangerous to release,' and the judge could order it held back.

Preemption of state laws: The bill would have created a unified federal framework, preempting the growing patchwork of state AI laws. For AI companies, this is a double-edged sword: one federal standard is simpler than 50 state standards, but a weak federal standard could preempt stronger state protections.

Why It Stalled: The Political Dynamics

The bill's delay resulted from three overlapping disagreements:

1. Anthropic vs. the framework. Anthropic, one of the leading frontier AI companies, pushed for stronger public disclosure requirements — arguing that AI companies should be required to share safety testing results, capability evaluations, and risk assessments publicly, not just with the government. Anthropic also opposed giving the Commerce Department ongoing injunction power, arguing it could be used to block legitimate AI development without adequate due process. Some Congressional staffers countered that Anthropic's preferred approach — heavy paperwork requirements — would burden startups while benefiting large, well-resourced AI companies like Anthropic itself.

2. Cantwell's concerns. Senate Commerce Committee Ranking Member Maria Cantwell (D-WA) had not yet agreed to the framework, and her support would have been necessary for the bill to advance through the committee with bipartisan backing. Cantwell's specific concerns haven't been fully detailed, but they reportedly include questions about whether the injunction mechanism is strong enough, whether the preemption provisions adequately protect stronger state laws, and whether the bill's definition of 'catastrophic risk' is appropriately scoped.

3. The open-weight complication. The bill landed in the middle of the intense open-weight vs. closed-source AI debate. Any regulation of frontier AI models necessarily raises the question of whether open-weight models — which anyone can download and modify — should be regulated differently than proprietary models accessed through APIs. The 77-company open letter backing open-weight AI, the Hugging Face security incident, and the 'Pacing the Frontier' petition all created cross-pressures that made the regulatory politics more complex.

What Happens Next: The Regulatory Trajectory

The bill's stall doesn't mean AI regulation is going away. Here's the most likely trajectory:

August-September 2026: The bill is revised to address Anthropic's and Cantwell's concerns. Key questions to be resolved: how strong should public disclosure requirements be? Should the injunction mechanism be modified? How should open-weight models be treated?

Fall 2026: The revised bill returns to the Senate Commerce Committee. If the core disagreements are resolved, markup and committee passage are possible. If not, the bill could stall indefinitely — a real risk given the complexity of the issues and the intensity of industry lobbying.

2027 and beyond: Even if the Thune bill doesn't pass, the pressure for federal AI regulation will continue building. The EU AI Act is in force. State AI laws are proliferating (California, New York, and others have active AI legislation). Major AI incidents (like the OpenAI breakout) create political pressure for action. The AI industry's own employees are calling for regulation. Something will pass eventually — the question is when, and in what form.

The state-level patchwork continues: While federal legislation stalls, states will continue passing their own AI laws. This creates compliance complexity for businesses operating across multiple states and strengthens the argument for a unified federal framework. The longer federal legislation takes, the more entrenched state-level regulation becomes.

What Businesses Should Do Now

1. Don't wait for US federal law — comply with the highest current standard. The EU AI Act is the most comprehensive AI regulation in force. If your AI systems or AI-powered products could be used in Europe, you likely need to comply. Building AI governance around the EU standard also positions you well for eventual US federal requirements, which are likely to be similar in structure if not identical in detail.

2. Monitor state-level AI laws. California, New York, Colorado, and other states have active AI legislation. If you operate in multiple states, you may already be subject to state AI requirements. Track these developments closely — state laws can take effect before federal legislation passes.

3. Build AI governance infrastructure now. Regardless of what specific legislation passes, certain requirements are nearly certain: documentation of AI systems and their capabilities, risk assessments for high-impact AI use cases, transparency about when and how AI is being used, and accountability mechanisms for AI decisions. Building these capabilities now is cheaper than scrambling to comply when legislation takes effect.

4. Pay attention to the open-weight regulatory debate. If your business uses open-weight AI models (Llama, DeepSeek, Qwen, etc.), the regulatory treatment of open-weight models will directly affect you. Restrictions on open-weight models could limit your access to free, self-hostable AI. A permissive approach could preserve access. This is the most consequential unresolved question in AI regulation, and its resolution will shape the AI landscape for years.

5. The self-regulatory window is closing. The 'Pacing the Frontier' petition, the OpenAI security incident, and the bipartisan interest in AI safety legislation all point in the same direction: the era of AI companies regulating themselves is ending. Businesses that have built their AI strategies assuming minimal regulation should begin planning for a more regulated environment. The specifics are uncertain, but the direction is clear.