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

Key Points:

• Anthropic has proposed a draft bill for AI safety regulation, suggesting government intervention to block AI model releases posing “foreseeable, imminent risks” such as potential deaths exceeding 50 people or property damage over a billion dollars.
• The bill outlines catastrophic risks stemming from scenarios like biological weapon creation, models evading developer controls, or cyberattacks, aligning with similar thresholds in California’s Senate Bill 53.
• Independent evaluators registered with the U.S. Commerce Department would assess AI models, with a 72-hour window to alert authorities if catastrophic risks are detected, after which a judge could halt model distribution; federal agencies cannot independently stop releases.
• The proposal emphasizes voluntary oversight by AI labs, with evaluators potentially funded by the labs but required to remain impartial, although concerns remain about the influence of payment on evaluations.
• Anthropic’s approach reflects broader industry trends favoring self-regulation and embedded third-party monitoring over mandatory government controls, despite ongoing fears about AI security and potential misuse.