Anthropic pitched US lawmakers on a draft AI safety bill. It lets the government block a model only if it risks over $1B in damages. It also applies if the model risks killing or seriously injuring over 50 people. The risk must come from a bio weapon, control evasion or a cyberattack. The bill requires testing, transparency and independent evaluation of frontier models. Evaluators must register with the Commerce Department. Labs can pay evaluators if payment is not tied to outcomes. Evaluators must report imminent risks to Commerce within 72 hours. Only a DC federal judge can order a model suspension. No other US agency can halt development or deployment. The bill aligns with Anthropic’s earlier Advanced AI Framework.
What it unlocks
Government power to block high-risk AI model releases or uses.
- $1B in damages threshold
- 50+ deaths or serious injuries threshold
- 72-hour reporting window
Sources