Anthropic has submitted a draft bill to US lawmakers aimed at restricting federal authority to block advanced AI models. According to the document, authorities would only be permitted to halt a model if it poses a predictable and immediate threat of damage exceeding $1 billion or risks causing death or serious injury to more than 50 people.
The proposal specifies that such threats must be linked to biological weapons, loss of model control, or cyberattacks. Under the draft, the Department of Commerce would be required to obtain a ruling from the US District Court for the District of Columbia before suspending any AI model, preventing unilateral executive action.
The bill introduces a framework for independent assessments, requiring evaluators to report threats to the Department of Commerce within 72 hours. While evaluators may receive compensation from companies, the pay must not be contingent on the assessment results. The project also mandates transparency reports and whistleblower protections.
Additionally, the proposal seeks to limit the authority of individual states to impose their own reporting and safety requirements, favoring a unified federal approach. This initiative aligns with the Advanced AI Framework that Anthropic has been promoting since June to establish safety standards for powerful AI systems.
The draft is currently under consideration as Congress debates various AI regulation strategies. While some lawmakers advocate for stricter oversight or development pauses, Anthropic's proposal offers a narrower scope for government intervention compared to broader pre-release screening initiatives.
