Artificial intelligence developer Anthropic has drafted a legislative proposal for a federal AI safety framework, suggesting that the deployment of advanced frontier models should only be restricted through formal judicial proceedings in extreme circumstances. Disclosed details indicate that the draft legislation specifically targets the most sophisticated AI systems rather than broad-based applications.
The proposed intervention threshold focuses strictly on three distinct high-risk areas: assisting in the creation or release of biological weapons, models escaping developer control, and large-scale cyberattacks facilitated or conducted by AI. Regulatory mechanisms would only trigger when risks originate from these specific scenarios. The risk standards outlined in the proposal bear similarities to California’s previously passed Senate Bill 53, which similarly defines catastrophic risks as incidents capable of causing mass casualties or severe economic damage.
Under the execution mechanism outlined in the draft, the legislation recommends establishing an independent assessment system registered and managed by the U.S. Department of Commerce. These independent evaluators and experts would be tasked with reviewing frontier AI models and their parent companies to determine if significant risks exist. The draft permits assessment agencies to be compensated by the AI companies being evaluated, provided that the payments do not have a direct correlation to the assessment findings. However, this financial arrangement has drawn skepticism from industry observers who question whether it could compromise the independence and objectivity of the evaluating bodies.
Should evaluators determine that a specific model presents an imminent catastrophic threat, they would be mandated to report their findings to the Department of Commerce within 72 hours of completing the assessment. Following this notification, the Commerce Department could petition a federal court in Washington, D.C., seeking a judicial order to suspend the release or deployment of the model.
Unlike several alternative regulatory frameworks under discussion, Anthropic’s draft explicitly stipulates that administrative agencies would not possess the authority to unilaterally halt an AI model. According to the proposed design, only a federal judge holds the power to issue a final suspension order, ensuring that government departments cannot bypass standard judicial procedures to prohibit the launch or operation of an AI system.
While Anthropic has not publicly released the complete final text of the bill, a company spokesperson confirmed its existence and noted that its core philosophy derives from the responsible scaling policy introduced earlier this year. That framework advocates for rigorous testing of the most powerful systems, increased transparency, and the introduction of independent third-party evaluations.
As the capabilities of artificial intelligence systems continue to expand, debates surrounding oversight and safety within the U.S. technology sector and regulatory bodies have intensified. Supporters argue that the framework provides a clear legal and procedural foundation for government intervention in high-risk scenarios. Critics, meanwhile, contend that high intervention thresholds and a reliance on industry-funded evaluations may prove insufficient to address future complexities in artificial intelligence risks.
The proposal remains in a preliminary discussion phase, has not been formally introduced to the U.S. Congress, and currently carries no legal force. Nevertheless, as one of the initial federal regulatory proposals initiated directly by a major artificial intelligence corporation, its provisions are expected to shape ongoing discussions regarding federal AI legislation.
