Anthropic Circulates Draft AI Bill Requiring Court Order to Block Models

Anthropic has circulated confidential draft legislation to U.S. lawmakers proposing strict federal limits on when regulators can block the release of advanced artificial intelligence models. First reported by Politico, the proposal would require federal authorities to meet high evidentiary bars in court before halting any model deployment.

Strict Damage and Casualty Thresholds

Under the draft measure, the federal government could only block or suspend a frontier model if it presents a foreseeable, imminent risk of killing or seriously injuring more than 50 people, or causing over $1 billion in economic or property damage. The intervention authority is restricted to three catastrophic scenarios: the creation or release of biological weapons, a model evading its developer's control, or cyberattacks.

Court Injunctions and Agency Limits

The draft bill designates the Department of Commerce as the sole federal agency permitted to seek a deployment halt, barring all other departments from intervening. Commerce officials would also be barred from unilateral emergency shutdowns. To block a model, the agency would need to obtain an injunction from a federal judge in the U.S. District Court for the District of Columbia under a preponderance of the evidence standard.

Evaluator Mandates and State Preemption

The proposal requires frontier developers to submit models to periodic evaluations by independent third parties registered with the Commerce Department. Developers would be allowed to pay these evaluators directly, provided the fees are not contingent on audit results. Evaluators would be required to alert the Commerce Department within 72 hours of discovering any qualifying imminent danger. Drawing framework language from California measures such as SB 53, the proposal would also override certain state-level AI regulations.

Criticism Over High Standards for Oversight

AI safety advocates have pushed back against the draft, arguing that casualty and monetary thresholds set an impractical bar while ignoring widespread harms such as algorithmic fraud. Sneha Revanur, founder and president of Encode Action, criticized the proposal as self-serving, stating that the government should retain emergency options without needing to prove whether 50 people or 49 will die. An Anthropic spokesperson defended the outreach, stating the draft reflects principles from the company's Advanced AI Framework, which it has shared with lawmakers since June 2026.