Anthropic pitches bill limiting AI intervention to catastrophic risks
The proposal gives one AI company influence over frontier safety rules.
Overview
Anthropic wants Congress to let the government stop an only when its danger is both foreseeable and imminent. The danger would have to risk more than $1 billion in damage, or kill or seriously injure more than 50 people. It would also need to involve a , a , or the . registered with the would have 72 hours to report such threats. could then ask a , to suspend the model. Other agencies could not act alone.
The plan would make government action clearer and harder to misuse, but serious harm below these limits could receive no federal response. Officials may struggle to estimate deaths or financial losses, confirm that evaluators are independent, and move quickly through court. Anthropic would also have an unusually large role in setting rules for its own industry.
Timeline
The proposal's enforcement mechanisms remain narrowly defined
The reported bill would limit government intervention to specified , require to notify the , and channel suspension requests through a federal court, leaving the practical scope of enforcement dependent on those thresholds and procedures.
Commerce would need court approval
The Commerce Department could seek an order from a to . The draft would bar other federal agencies, departments, and officers from independently stopping development, deployment, or internal use.
Registered evaluators receive reporting duty
registered with the would periodically assess and models. If they identified an imminent risk meeting the bill’s criteria, they would have to notify Commerce within 72 hours of completing the assessment.
Three risk categories qualify
The qualifying risks would have to involve the creation or release of a , a , or a . Limiting intervention to these categories would narrow the circumstances in which the government could halt a model.
Bill sets a narrow intervention threshold
The draft would allow intervention only for a causing more than $1 billion in damages, or the death or of more than 50 people. could make more predictable, but potentially leave severe harms below the line unaddressed.
Anthropic pitches draft bill to lawmakers
Anthropic representatives presented lawmakers with a draft AI safety bill, according to reporting by Politico. The proposal sought to turn the company’s broader into specific , procedures, and .
Anthropic publishes catastrophic-risk framework
Anthropic publicly proposed a legal framework for addressing from . The framework envisioned to block or deter , focused on the largest models and companies with substantial AI research or revenue.