Overview
- A privately circulated draft bill from Anthropic was leaked and publicly confirmed by the company as its Advanced AI Framework proposal.
- The draft defines a “foreseeable, imminent” catastrophic risk with numeric triggers such as more than 50 potential deaths or over $1 billion in property damage to allow blocking a model’s release.
- It would rely on Commerce-registered “independent evaluators” to flag imminent risks within 72 hours, require Commerce to seek a D.C. court order to halt release, and bar other federal agencies from acting on their own.
- The proposal permits evaluators to be paid by AI labs if payments are not contingent on results, a feature safety advocates say risks conflicts of interest and weakens impartial oversight.
- Advocates and rivals have pushed back, arguing lawmakers—not a single company—should write AI rules, and the draft has intensified debates over federal preemption of state laws and the proper design of enforcement and oversight.