Anthropic Proposes AI Safety Bill Limiting Government Power Over Advanced AI Models

Anthropic Proposes Strict Thresholds for Government Intervention on AI

Anthropic has presented US lawmakers with a confidential draft AI safety bill that would establish strict limits on when federal authorities can halt advanced AI models from reaching users1

1

. The proposal, obtained by POLITICO, would allow government intervention only when a model presents a foreseeable, imminent threat involving more than $1 billion in damages or the deaths or serious injuries of over 50 people2

2

. This represents a significant attempt by Anthropic to shape AI regulation while restricting the circumstances that could trigger government action against frontier AI safety systems.

The bill limits qualifying threats to three specific scenarios: creating or releasing biological weapons, models escaping their developers' controls, or cyberattacks3

3

. Other forms of harm would fall outside this specific intervention framework. An Anthropic spokesperson stated that the legislation reflects the company's Advanced AI Framework, which it has promoted among federal and state lawmakers since June, calling them "sensible policies we believe Congress should act on"1

1

.

Independent Evaluators and Mandatory Reporting Requirements

Under the proposed AI regulation, independent evaluators would register with the Department of Commerce and periodically assess frontier AI developers and their models5

5

. The bill would allow labs to pay these evaluators, provided their compensation does not depend on their findings4

4

. If an assessment identifies an imminent threat covered by the bill, evaluators must notify Commerce within 72 hours of completing their review2

2

.

This evaluation framework aims to balance oversight with industry flexibility. The 72-hour reporting window for catastrophic threats ensures rapid response while maintaining independent scrutiny of advanced AI models. The provision allowing developer payment of evaluators, though potentially controversial, includes safeguards against conflicts of interest by prohibiting outcome-based compensation.

Judicial Approval Required for Model Suspensions

The draft bill would prevent the Department of Commerce from independently shutting down a model1

1

. Instead, the department would need to seek an order from a federal judge in the District Court for the District of Columbia3

3

. The proposal explicitly bars other federal departments, agencies, and officials from suspending a model's development, deployment, or internal use2

2

. This restriction could reduce the risk of sudden executive action against AI developers.

Requiring judicial approval represents a significant check on government intervention. The issue has particular relevance following US export controls that reportedly sidelined two advanced Anthropic models over the summer1

1

. That episode raised concerns across the technology industry about government officials restricting access to powerful systems without clear legal frameworks.

Federal Preemption and State Law Conflicts

The bill would preempt certain state requirements covering public risk disclosures, independent evaluations, and incident reporting1

1

. This provision could reignite disputes over whether Washington should establish a national standard or preserve stronger state protections. The language tracks California's AI laws, including SB 53, signed into law last year4

4

. Democrats have resisted such federal preemption proposals in the past.

This federal-state tension highlights a critical question in AI regulation: whether a single national framework can adequately protect the public without undermining state-level innovations in safety standards. Some lawmakers and safety advocates worry that federal preemption could weaken existing state AI laws that provide stronger protections.

Competing Legislative Proposals and Industry Debate

Anthropic's draft enters a crowded legislative debate over how the United States should regulate frontier AI. Representatives Lori Trahan and Jay Obernolte have advanced the FRONTIER Act, which would authorize government intervention against risky models and override some state laws1

1

. In the Senate, Majority Leader John Thune has worked with Senators Amy Klobuchar and Ted Cruz on legislation targeting catastrophic threats4

4

. Senator Maria Cantwell and some safety advocates want a stronger federal review process.

Other US lawmakers have pushed for more sweeping restrictions. Representative Greg Casar and Senator Bernie Sanders have called for pausing advanced AI development and banning superintelligence1

1

. Anthropic's proposal takes a narrower approach, requiring frontier developers to establish safety frameworks, publish transparency reports for new models, protect whistleblowers, and report critical incidents to Commerce1

1

.

The broader industry remains divided on safety priorities. Last month, Anthropic researcher Evan Hubinger put the odds of AI wiping out humanity within a decade above 10%, amid ongoing debates about imminent risks versus long-term existential threats4

4

. Whether lawmakers accept Anthropic's proposed limits will help determine how much authority the federal government gains over advanced AI models and whether regulators receive excessive discretion in limiting federal power over emerging technologies.