Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling

The author argues that Anthropic’s circulated draft federal AI safety bill and its public response miss the central problems: the text privileges industry interests and accepts compromises that are unacceptable given the stakes. Four core objections are highlighted. First, preemption language would strip states of critical regulatory tools rather than substitute a clearly stronger national standard. Second, model‑blocking provisions set an impractically high legal bar for government intervention; the author insists emergency takedowns must be workable in practice (citing a §13‑style restraining power and past incidents like Fable 5) rather than delayed by heavy evidentiary burdens. Third, risk thresholds were borrowed from California law and remain too high in practice (the author notes real examples where harms short of billion‑dollar losses or mass fatalities still warrant reporting and action). Fourth, proposed updates - like allowing courts to halt internal use and requiring breach reporting - are improvements but insufficient; they make the draft a modestly altered FRONTIER‑style framework that is weaker overall.

The piece contends the draft reads like regulatory capture: structured to protect Anthropic’s deployment plans, preserve business flexibility, and constrain state action, even while claiming to be “the strongest bill on the table.” The author points to shifting political winds - companies pulling back political spending, growing pro‑safety coalitions, and state lawsuits seeking to halt advanced testing - and urges honesty and bolder standards (citing Sanders’ proposal as stronger). Concerns about evaluator independence and “evaluator shopping” are raised as additional technical flaws that make the draft troubling rather than constructive.