Leaked Anthropic draft trades state safeguards for a weaker federal brake

A discussion draft for a US federal law has leaked. It is attributed to Anthropic and was reportedly circulated to lawmakers by the company. It has not been introduced as a bill, and Anthropic has not publicly confirmed it as a final position, so the leaked text must be separated from adopted company policy.

A federal regime would take powers away from states

According to the published text and legal commentary, the draft would preempt state rules covering risk disclosures, external evaluations and incident reporting. Critics say that would reach California's SB 53 and New York's RAISE Act. The draft also contains continuing audits, risk reports, whistleblower protections and security duties.

The contentious part combines preemption with a high threshold for federal intervention. Restricting model development or deployment would require an imminent risk of catastrophic harm. Analyses of the leaked language describe that as more than 50 deaths or over $1 billion in damage, alongside a requirement to consider less restrictive alternatives.

The fight is over who may pull the emergency brake

A single federal framework could spare labs from overlapping rules across 50 states. But if federal law weakens enforcement while preventing states from adopting stronger safeguards, uniformity becomes a ceiling. The result is not merely less paperwork for companies; it is fewer institutions able to intervene.

That sits uneasily with Anthropic's own June framework. The company publicly argued that Congress should not preempt state law without a rigorous federal regime at least as strong as its proposed measures, and that any preemption should be narrow. Critics say the leaked draft offers broader insulation from states in exchange for weaker federal powers.

A leak and hostile commentary do not yet equal settled policy

The sharpest claims come from critics of the proposal, not an official Anthropic announcement. A discussion draft can contain errors and change during negotiation. Without a public response, it remains unclear who approved the text, how actively Anthropic supports it and which clauses are bargaining positions.

Dismissing it merely because it is provisional would also be a mistake. Federal proposals tied to preemption can lock in a compromise for years. The exact powers of courts, auditors and agencies matter more than the general label of an AI safety bill.

Anthropic's response and a congressional sponsor will reveal the intent

The first signal is whether Anthropic confirms authorship and explains the gap from its June framework. The second is any congressional sponsor and changes to preemption, emergency restrictions, auditor licensing and agency rulemaking.

Until someone introduces the text, it is a policy proposal rather than law. Yet the negotiating baseline is being set now: a federal floor that states may build on, or a low ceiling that blocks them from going further.

Lilith's verdict

When a lab designs a brake that makes states release the lever and lets Washington pull only near the wall, it is fair to ask whose hands the mechanism protects.