Anthropic pitches Congress a draft AI safety bill

By Splitfeed EditorialLast updated Oct 19, 2026, 17:00 UTC

On October 16, 2026, Politico reported that Anthropic had privately presented a draft AI safety bill to US lawmakers before the details became public. The proposal would give the Commerce Department exclusive authority to block an AI model from deployment, but only when a specific harm threshold has been crossed: at least 50 deaths or at least $1 billion in economic damages. Anthropic made the pitch quietly, and the bill's contents became public through Politico's reporting rather than any announcement from the company.

The thresholds are the central point of contention. By setting a high floor for government action, the draft would effectively insulate almost all AI deployments from federal interference, reserving intervention for catastrophic outlier events. That framing cuts against Anthropic's longstanding public identity as a safety-first company founded precisely because its founders believed frontier AI posed serious risks. The bill also consolidates regulatory authority in a single agency, the Commerce Department, leaving out bodies like the FTC or sector-specific regulators that might otherwise have a claim to oversight. Whether that concentration of authority makes enforcement more coherent or more capturable is genuinely disputed.

The pro-AI camp, as reflected in a post from the account Ea-nāṣir, frames the bill as a measured attempt to limit unnecessary government blocking of model releases, treating the high threshold as a reasonable guard against regulatory overreach. The anti-AI camp is sharply critical on multiple fronts. The account eclipticplane and revolvingthrow both mocked the specific numbers through trolley-problem analogies, arguing the thresholds create perverse incentives where severe but sub-threshold harms go unaddressed. Yizahi raised the issue of IP theft, arguing that the scale of alleged training-data appropriation already dwarfs the $1 billion damage threshold. The account createful framed the bill as regulatory capture designed to disadvantage open-source models by labeling them dangerous. The middle ground, represented by arw0n, notes that Anthropic finds itself in an unwinnable public relations position while larger rivals like Google and Microsoft stay quiet on legislation. The account ponder offered an alternative framework, suggesting that oversight modeled on licensed professional engineering sign-offs might be more credible than industry-adjacent panels deciding which models reach the public.

The immediate thing to watch is whether the draft bill gains a sponsor in Congress and advances to a formal hearing, which would force Anthropic to defend the specific thresholds in public testimony and give competing voices, including open-source advocates and civil society critics, a formal venue to respond.