Artificial intelligence lab Anthropic has circulated a draft legislative proposal outlining how the federal government might regulate catastrophic risks associated with advanced machine learning models. The proposed bill establishes strict thresholds for intervention, suggesting that a model's public distribution should only face blocks if it poses a foreseeable, imminent risk of resulting in the deaths of more than 50 people or causing over one billion dollars in property damage. These thresholds echo metrics previously utilized in recent California legislation, such as Senate Bill 53, which similarly classified those specific outcomes as catastrophic events.
Under the terms of the drafted framework, the assessment of a model would rely on independent evaluators registered with the U.S. Commerce Department. These third-party examiners could potentially receive compensation directly from the artificial intelligence labs, provided the payment structure remains independent of evaluation outcomes. Critics note that questions remain regarding how such a payment system could completely eliminate conflicts of interest. Once an evaluator identifies an imminent threat stemming from biological weapon creation, runaway model autonomy, or cyberattacks, they would face a 72-hour window to notify the Commerce Department.
The legislation introduces a specific procedural hurdle for federal oversight by explicitly stripping federal agencies of the unilateral authority to halt an artificial intelligence model's distribution. Instead, the Commerce Department would be required to petition a Washington, D.C. judge to legally block the release. Anthropic representatives confirmed the existence of the draft text, noting it stems from the company's Advanced AI Framework. While tech companies have frequently warned about existential threats tied to artificial intelligence, the industry has historically pushed back against mandatory government control, favoring self-govergulation and independent monitoring frameworks instead.
Comments