In brief
The proposal would use independent assessments and federal risk thresholds to govern restrictions on AI models, RBC-Ukraine reported, citing Politico.
Anthropic has presented US lawmakers with a legislative proposal to restrict the release or use of AI models posing catastrophic risks, RBC-Ukraine reported on October 21, citing Politico. The report did not specify when the proposal was presented or report its adoption.
Under the proposal, intervention would require a foreseeable and imminent risk. Proposed criteria include potential deaths or serious injuries affecting more than 50 people, property damage exceeding $1 billion, and threats involving biological weapons, large-scale cyberattacks or developers losing control of a model. These describe potential consequences, rather than harm already recorded.
Independent evaluators accredited by the US Commerce Department would assess safety and notify the department within 72 hours of identifying a catastrophic risk. The department would then have to seek action in a federal court in Washington. Federal agencies and officials would be barred from independently blocking model development or release, according to the report.
The proposal would also replace separate state laws with a single federal standard. RBC-Ukraine reported concerns that this could weaken stricter state protections and that allowing AI laboratories to pay evaluators could create conflicts of interest. Such payments would be permitted only if fees did not depend on the assessment's outcome.