Anthropic's reported federal AI safety proposal is facing criticism from both safety advocates and a competitor. The supplied account describes planned revisions by federal-affairs lead Anthony Cimino, criticism from Encode AI's Nathan Calvin and a governance statement by OpenAI's Chris Lehane.
Semafor's October 20 reporting confirms the broader dispute: a circulated draft became public, Anthropic discussed changes, and critics questioned its treatment of state powers and catastrophic risk. This remains a proposal debate, rather than a new law taking effect.
Illustration from the supplied AREX/BAAI Hub page; not an image of a legislative document.
The source says Politico disclosed the draft on October 16. It describes a legislative version of Anthropic's Advanced AI Framework, with a federal process for restricting dangerous models and requirements for outside evaluation.
Several reported details are especially consequential:
| Issue | Description in the supplied reporting | Evidence limit |
|---|---|---|
| Risk threshold | Imminent risk of more than 50 deaths or serious injuries, or more than $1 billion in damage. | The complete clause and definitions were not checked against the draft text. |
| Restriction procedure | A government request requiring approval by a federal judge in Washington, DC. | Procedural descriptions come from reporting, rather than a verified final bill. |
| Evaluator oversight | Commerce registration, independence despite developer payment, and notification within 72 hours after an evaluation finding the specified imminent risk. | The trigger is not described as a blanket 72-hour duty for every evaluation. |
| State-law treatment | Federal rules would displace some state authority. | The exact scope is disputed and revisions have been proposed. |
The reported thresholds concern a proposed intervention mechanism. They do not mean that smaller harms are lawful, acceptable or exempt from every other legal obligation.
Planned Revisions to Preemption and Risk Thresholds
Preemption concerns when federal law displaces state law. Here, the disagreement is about which AI safety powers would remain with states after a federal framework was adopted.
Anthropic's public framework says Congress should displace state safety rules only through sufficiently strong federal legislation, and describes the intended displacement as narrow. It says states should retain authority over matters outside the federal framework's specific safety functions, including child safety and consumer protection.
The supplied account says Cimino intends to clarify that states retain important roles involving independent evaluators and oversight of laboratories' ability to stop models. His reported explanation also acknowledges that some lawmakers want legislation that resolves overlapping regulatory authority.
The source attributes the proposed risk threshold to an existing California standard, and says Cimino considers it worth reassessing. Without the complete proposed text and the exact state-law provision side by side, that account should not be taken as proof that the two operate identically.
Reporting and Evaluator Independence
According to the supplied account, proposed updates also address dangerous internal use, breaches of containment and access to outside systems. It describes further work on incident-reporting obligations, evaluator standards and preventing “evaluator shopping,” where developers seek the most favorable reviewer.
Those concerns align with the public framework's broader requirement for qualified independent evaluation. Its stated agenda includes published safety summaries, regular risk reporting and stronger security programs for frontier developers.
The framework also sets a proposed scope: models trained with more than 10²⁵ FLOPs, developed by companies above specified AI revenue or research-spending thresholds. Those framework criteria should not be silently substituted for the scope of the separate leaked legislative draft.
The Case for Emergency Powers With Safeguards
The supplied account describes Cimino defending a court-supervised mechanism as a balance between rapid intervention and preventing abuse. The important policy tension is how an authority can act before catastrophic harm occurs while remaining accountable for a mistaken restriction.
Anthropic's public framework calls for government authority to block or deter dangerous deployments, together with safeguards against excessive regulatory power. It also discusses transparency, model security and broader resilience measures for biological, cyber and loss-of-control risks.
That public framework establishes the company's policy position. It does not settle whether the leaked draft's legal language implements that position effectively.
The source says proposals about controlling the pace of frontier development were still being explored, rather than already fully translated into legislative text. A proposal to study future pacing policy is different from a current, enforceable limit on model development.
Critics Question the Balance and the Rulemaking Process
The source attributes two central objections to Nathan Calvin: that the preemption language is too broad and that the federal intervention standard is too difficult to satisfy. It recounts concerns about state-court powers, a demanding showing of imminent harm and the court's treatment of the government's analysis.
Build a showcase site and grow leads in minutes
Describe your idea once, and We0 AI can generate a showcase site, pages, and CMS, then help you attract customers and traffic after launch.
One complete project generation for free registration
Best for trying one complete generation flow and seeing a first project draft quickly.
These are a critic's interpretation of the draft. Because the full document was not available for this coverage, they should not be presented as adjudicated conclusions about its legal effect.
Lehane's separate position, as reported by Semafor, supports national safety standards while emphasizing democratic decision-making. His statement argues that elected officials should lead and that no single company, including OpenAI, should write the rules alone.
The supplied account connects that statement with support for a bipartisan FRONTIER Act. That reference does not establish that Anthropic's circulated draft is the same bill, that either has passed or that all participants agree on the details.
Semafor's comparison of the companies' Washington strategies is editorial analysis. It should remain distinct from the narrower facts that a draft circulated, criticism followed and changes were discussed. The source's references to other government-company disputes do not determine the legal merits of these provisions.
Changes Remain Prospective
The supplied source says the announced revisions had not yet appeared as a complete public replacement text. Accordingly, neither the intended preemption changes nor a revised threshold can be treated as settled wording.
Its description of congressional timing also does not establish that no federal law touches AI. Existing laws can apply to AI-related conduct even in the absence of a comprehensive frontier-model safety statute.
The next useful documents would be the revised draft, a clearly identified introduced bill, its definitions and its legislative status. Those would allow a direct comparison of state powers, emergency intervention standards, evaluator duties and any development-pacing proposal.
FAQ
Is Anthropic's reported AI safety draft already law?
The supplied reporting describes a circulated proposal. It does not establish enactment or a current obligation created by that proposal.
What does state-law preemption mean here?
It concerns federal rules displacing some state authority. The disputed issue is how broad that displacement would be and which safety powers states would retain.
Does the reported threshold permit harm below that level?
No such conclusion follows from the reporting. A threshold for one proposed intervention power is different from the reach of other laws and duties.
Are the planned changes already written into a final bill?
No complete revised legislative text was obtained for this coverage. Statements about intended changes should remain separate from enacted or finalized wording.
Why is independent evaluation controversial?
Reviewers need enough access and expertise to assess risks while remaining independent of the developers paying for their work. The source discusses concerns that developers could seek favorable evaluators instead of rigorous scrutiny.
Did OpenAI say companies should have no role in policy?
The statement reported here calls for elected officials and democratic processes to lead. Its objection is to any one company writing the rules alone, rather than a ban on industry input.
Related Tools
- Congress.gov: The official service for checking introduced federal bills, their text and legislative status.
- Federal Register: The official publication for checking federal agency rulemaking and notices.
Related Links
- Supplied AREX article on BAAI Hub: The aggregation that prompted this coverage, including its source limitations.
- Anthropic's public policy framework: The company's published position on frontier governance and state-law preemption.
- Semafor's October 20 report: Reporting on the leaked draft and the resulting disagreement.
- Anthony Cimino's source-listed response: The social statement cited by AREX; direct retrieval was blocked during verification.
- Nathan Calvin's source-listed criticism: The critique referenced in the supplied account.
- Chris Lehane's governance statement: The LinkedIn statement linked by both AREX and Semafor.
Summary
The dispute centers on the relationship between a federal intervention process and existing state powers, as well as whether reviewers and regulators could act effectively. Anthropic has discussed changes, while critics question the draft's balance and OpenAI emphasizes elected leadership.
A circulated proposal and a promise to revise it are not a new law; the actual revised text and legislative status remain essential to assessing its effects.
