Encode General Counsel Nathan Calvin and Charlie Bullock, Senior Policy Director at the Institute for Law & AI, are challenging the intervention powers in Anthropic's circulating AI bill. In their own October 16 and 17 posts, they warn of weaknesses in enforcing safety rules. Bullock also credits improvements in external evaluators' access.
Calvin argues that the draft could displace state protections while making intervention against dangerous models harder. He objects particularly to broad federal preemption combined with demanding conditions for court orders suspending model development or use. That is his legal assessment of a proposal, not an established change in current law.
Bullock also considers rulemaking and emergency intervention powers weaker than in the FRONTIER Act. But limits could protect against an administration abusing those powers, he argues. He welcomes continuing evaluator engagements and information access. In a subsequent post, he calls for government licensing of evaluators rather than relying on registration and labs' selection.
New York Assemblymember Alex Bores calls for stronger safeguards after recent AI incidents. Nat Purser proposes enforceable corrective action after adverse evaluations, evaluator selection and payment independent of labs, and adaptable minimum standards. These are individual proposals, not an announced joint initiative.
Zvi Mowshowitz, Nate Soares, Sydney von Arx and Kelsey Piper also criticize the balance between protection from dangerous AI and protection of developers from government intervention. Mowshowitz explicitly says he had not read the entire bill. Their contributions contain personal judgments. Piper's invented dialogue between company departments is satire, not an insider account.
The legal interpretation is contested. A counterargument published under the name ModelMirror says an explicit statutory route could make federal intervention easier. It points to a clause preserving authority under other federal laws. In the publicly linked discussion copy, the operative preemption provision also names only three of four previously listed areas, omitting model suspension. Whether that omission is intentional remains unclear. ModelMirror assesses particular provisions rather than endorsing the draft overall.
The circulating copy is labeled as a discussion draft last edited September29. Anthropic told Reuters that its proposals reflect its Advanced AI Framework. The company described them as a way to prevent dangerous deployments of powerful models and require testing, transparency and independent evaluation. The public responses offer contrasting judgments about how to connect those evaluations with effective powers to intervene.