Anthropic's Draft AI Safety Bill

Under Anthropic's draft, the government cannot unilaterally stop a frontier model - the Commerce Department would have to petition a federal judge in the U.S. District Court for the District of Columbia, and a suspension can only be ordered if a model poses a "foreseeable, imminent risk" of more than $1 billion in damages or of killing or seriously injuring more than 50 people, and only if that harm flows from bio-weapons creation, loss of model control, or a cyberattack ^[1][2]. No other federal agency may act independently. Independent evaluators - whom labs are allowed to pay, as long as payment isn't tied to the outcome - would register with Commerce and have just 72 hours to report an "imminent enumerated risk" once they find one ^[1].

The bill would preempt disclosure, third-party evaluation, and incident-reporting laws in states like California ^[4]- including provisions that mirror SB 53, the very law Anthropic publicly endorsed in September 2025, arguing that "powerful AI advancements won't wait for consensus in Washington" ^[3]. Policy commentator Nathan Calvin argues the preemption clauses sweep far beyond disclosure rules, potentially barring states from enforcing laws that suspend AI development or deployment even to protect citizens from imminent catastrophic harm ^[4]. That reversal looks sharper in hindsight: Anthropic's own June 2025 "Advanced AI Framework" said federal preemption should happen only if Congress passed a regime at least as strong as Anthropic's own proposal ^[5]- and critics including Alex Bores say the leaked bill is weaker than that framework, and in places weaker than the state laws it would override ^[4].

Not every critique is pure outrage. AI policy analyst Charlie Bullock calls the underlying idea "pretty reasonable" in principle, but flags that the bill has no government licensing requirement for the independent auditors it leans on - creating room for a race-to-the-bottom among evaluators competing for lab business ^[4]. Zvi Mowshowitz is far less charitable, arguing Anthropic is "protecting their business interests and guarding against their political risk, at the expense of safety considerations" ^[4]. The gap between those two reads - a workable-but-flawed framework versus a self-serving one - is itself the story: both camps agree the auditor system is unlicensed and the thresholds are narrow, they just disagree about whether that's negligence or design.

Public reaction turned hostile fast once the thresholds became legible outside policy circles. The most-shared framing on X distilled the bill into a single, mockable line - that Anthropic had effectively drafted itself room to cause dozens of deaths and nearly a billion dollars in damage before anyone could even request a suspension - a reading that traveled far faster than the bill's more technical provisions. On Reddit, the dominant theme was regulatory capture: commenters called the thresholds arbitrary and framed industry support for regulation as a way to raise barriers to entry for smaller competitors, while a minority pushed back that some friction is an unavoidable part of real safety rules. A smaller pro-Anthropic thread on X argued the company still deserves credit relative to rivals, but that view was a distinct minority against the otherwise critical spread of the Politico scoop.

Congress already has a rival, bipartisan alternative on the table: Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA) introduced the FRONTIER Act (H.R. 9925) in July 2026, months before Anthropic's pitch became public ^[6]. Critics say that bill carries stronger rulemaking and emergency-shutdown authority than Anthropic's draft, raising the real question for the months ahead: whether lawmakers treat Anthropic's proposal as a floor for negotiation or as evidence the industry would rather write its own rules than accept someone else's. Either way, the unlicensed-auditor gap flagged by Bullock ^[4]and the preemption fight over SB 53 ^[4]are likely to be the two provisions fought over hardest if this bill - or anything like it - reaches a formal markup.