Anthropic is revising its draft federal AI safety legislation following backlash from security experts and US lawmakers. The initial proposal, leaked by Politico, faced criticism for being less stringent than the bipartisan Frontier Act and potentially hindering the restriction of dangerous models.
The draft suggested that authorities could only ban a model if risks were immediate and predictable, with potential damages exceeding $1 billion or causing death or serious injury to over 50 people. Under this framework, the Department of Commerce would need federal court approval to act, while other agencies would be barred from suspending model operations independently.
Concerns were also raised regarding the limitation of state-level regulatory powers and the lack of standardized licensing for independent auditors. Critics, including California Senator Scott Wiener, argued that the proposal fails to ensure public safety and risks allowing developers to select lenient auditors.
Anthony Cimino, Anthropic's head of federal affairs, acknowledged on October 18 that the language regarding federal preemption was imprecise. The company plans to re-evaluate risk thresholds, refine the model-blocking process, and introduce stricter requirements for independent evaluators to prevent conflicts of interest.
The company also faces pressure from the other side of the spectrum, as former White House AI official David Saks warned that lower thresholds for advanced models could stifle startups in Silicon Valley. Anthropic has yet to release a final version of the revised legislative proposal.
