BusinessClaude

Anthropic proposes blocking AI that could kill 50+

Original title: Anthropic pitches AI safety bill that may block a model if 50+ people could be killed

Plain language

What it is: Anthropic has proposed a draft bill on AI safety to the U.S. Congress. The bill suggests blocking the release of AI models that pose extreme risks to human life or property.
How it works: Independent evaluators would check the models for risks. If they find a major threat, they report it to the government, which must ask a judge to stop the release.
Where it is available: The proposed bill is aimed at federal regulation in the United States.
Example: An AI model could be blocked if there is an imminent risk of it being used to create a biological weapon.

What it is: Anthropic has drafted legislation that establishes legal thresholds for blocking AI model distribution based on catastrophic risk.
How it works: Registered independent evaluators assess models for risks like cyberattacks or loss of control and must report findings within 72 hours. The Commerce Department must then petition a federal judge to halt the release, as federal agencies are barred from blocking models unilaterally.
Where it is available: The draft bill is intended for federal AI regulation and distribution policies in the United States.
Example: A model could be blocked from release if evaluators find it can evade its developer's security controls.

The headline and explanation on this page are based on TechCrunch's post. The full details are in the original.

Anthropic pitches AI safety bill that may block a model if 50+ people could be killed