Overview
- The confidential draft, disclosed by a Politico report on Friday, is being circulated privately to lawmakers and reflects Anthropic’s June Advanced AI Framework rather than an introduced bill.
- It lets federal authorities seek a model suspension only when a foreseeable, imminent risk would cause more than $1 billion in damages or kill or seriously injure over 50 people and only for three risks: biological weapons, models evading developer controls, or cyberattacks.
- The proposal defines covered ‘frontier’ developers by a technical line — training runs that exceed 10^25 floating‑point operations (FLOPs, a measure of compute used in model training) — plus a financial test tied to large AI revenue or R&D spending.
- It would require registered independent evaluators to run periodic tests and report any qualifying ‘imminent enumerated risk’ to the Commerce Department within 72 hours, and would bar other federal agencies from unilaterally suspending a model while directing Commerce to seek a D.C. federal court order for halts.
- Supporters call the draft a pragmatic, predictable framework while critics say its high damage/death thresholds, fixed FLOPs trigger, and federal preemption could weaken oversight, favor big labs, and leave many harms outside federal reach as Congress weighs competing bills.