Anthropic's draft AI bill: what it would mean for you

Anthropic has privately shown US lawmakers a draft bill that would set a high bar before the government could stop an AI model from being released. It is a proposal, not law, and its details come from second-hand reports.

According to TechSpot, which cites reporting that originated with Politico, Anthropic presented the draft to lawmakers. We have not seen the bill's text or its sponsors, so the details below come from summaries of the draft.

Under the draft, a model could be blocked only if it poses a foreseeable, imminent risk of either of these:

Only three kinds of risk would count: biological weapons, models escaping their developers' control, and cyberattacks.

The process would run through the courts. The Commerce Department would have to make its case in federal court, and a judge in the US District Court for DC would decide whether to stop a release.

One report says the draft would expressly stop any federal agency from halting distribution of a model on its own. One account says it would also limit state interventions to the most severe scenarios.

An Anthropic spokesperson described the ideas as "sensible policies we believe Congress should act on."

Nothing changes today. The draft is not law, and the reports do not say whether any lawmaker has taken it up.

If it did pass in this form, the AI tools you use would be much less likely to be pulled by a regulator. Action would be reserved for the three risk types above, and only once the damage or casualty threshold was in sight.

Critics argue that requiring a court decision makes blocking a model so hard procedurally that agencies might not even try. That is their reading of the draft, not an established outcome.

The notes do not say when, or whether, the draft will be formally introduced. The bill's text and sponsors have not been seen, so the details above may change once the actual wording is public.