AI Safety Blogger Zvi: Anthropic Doesn't Understand Why Its Draft Bill Was Unacceptable and Troubling
AI safety blogger Zvi Mowshowitz analyzes the response from Anthropic's federal affairs lead Anthony Cimino to the leaked US federal AI safety draft bill, arguing that the response addressed only peripheral details while missing the core problem: the bill's structure makes AI regulation harder rather than easier, and Anthropic's 'mistakes' systematically favor its own interests, contradicting the safety-organization identity its co-founders publicly articulated in 2023.
Anthony Cimino, who leads Anthropic's federal affairs team, responded over the weekend to a leaked US federal AI safety draft bill that had been shared with the press, addressing four points of criticism one by one: the scope of federal preemption over state law, the evidentiary threshold for model blocking, risk trigger thresholds, and updates following summer incidents. Zvi Mowshowitz—a long-time AI safety blogger and author of Don't Worry About the Vase—argues that this response precisely reveals Anthropic's failure to understand what was centrally wrong.
Zvi's core judgment is that Cimino treats the draft's problems as fixable details, but the real issue is structural. The bill vests all power to address catastrophic and existential AI risks exclusively in the US federal government, with a preemption clause stronger than the existing FRONTIER Act, while simultaneously setting an extremely high threshold for federal authorities to block a model. In Zvi's words, 'you cannot have it both ways'—where preemption is needed (catastrophic risk), it demands federal monopoly power, but in areas where preemption could actually unify fragmented state laws (consumer protections, deepfakes), the bill proposes no preemption at all.
On model blocking, Cimino draws an analogy to emergency handling of occupational safety or chemical risks. Zvi rejects the analogy: occupational safety and chemical risks involve ordinary amounts of danger where harms can be weighed against benefits; frontier AI risks can involve mistakes that cannot be undone, including permanent loss of control over the future or very large catastrophic dangers. He argues that model blocking obviously needs a 'take it down first and we argue about it later' mechanism, and while the draft contains that layer of meaning (§13 allows the US Secretary of Commerce to petition a district court for a temporary restraining order), it is packaged behind a high evidentiary bar presented as a 'meaningful standard.'