Anthropic has proposed a federal framework for assessing powerful AI systems, with government intervention limited to cases meeting severe harm thresholds. The reported details would create formal oversight while also restricting when officials could halt a model’s development or release.
Anthropic Watch analysis
What happened
The draft would allow intervention only where a model threatens more than $1 billion in financial damage, or death or serious injury to more than 50 people. The proposal identifies biological weapons, AI systems operating beyond their developers’ control and cyberattacks as possible threat categories.
It would also require independent evaluators registered with the US Commerce Department to assess developers and their most capable models. Evaluators who find qualifying serious risks would have 72 hours to notify the department. Commerce could seek judicial intervention, but a federal judge would decide whether a model should be suspended. Other agencies would be barred from independently stopping development, deployment or internal operation under this framework.
These details come from PYMNTS’ account of Reuters’ reporting on Politico’s original report. The draft is a proposal, not enacted law, and the reporting does not establish whether lawmakers intend to introduce it formally. Read the PYMNTS report.
Why it matters
This is more than a general call for AI safety rules. The proposed thresholds would reserve government intervention for extreme harms, while evaluator reporting and court approval would create a defined path for acting on them. At the same time, limiting other agencies’ authority could constrain how government responds outside that path.
The arrangement also raises a live question about independence: developers could pay for assessments, provided payment did not depend on evaluators’ conclusions.
Our read
The figures make the proposal easier to scrutinise, but they do not settle whether its safeguards are strong enough or its intervention bar too high. The draft’s wording and how evaluators would be appointed matter more than the reassuring phrase “independent assessment”. A rulebook can have a very precise threshold and still leave the biggest argument just above it.
What to watch
- Whether Anthropic or lawmakers release the draft text.
- Whether any lawmakers formally introduce the proposal.
- How the evaluator process would protect independence and handle urgent risks.
Discussion spark: Should AI companies help set a high legal threshold for government intervention in their models, or should that power remain broader and more flexible?
Sources and evidence
- Anthropic Proposes AI Safety Bill Giving US Government Power to Block High-Risk Models – PYMNTS.com (18 October 2026, 17:13 UTC)
- Source update (18 October 2026, 17:13 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.