Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling

Anthropic finally put out a response through Anthony Cimino, its federal affairs lead, after the draft federal AI safety bill leaked over the weekend. The answer tries to frame the document as a translation of its Advanced AI Framework into legislative language, shared with a small set of congressional offices. It also says the company is still refining pieces that are too new to turn cleanly into policy text.

That may explain the draft. It doesn’t explain why the draft landed so badly. The core criticism is not about a few technical edits. It’s that Anthropic circulated what looks like a compromise-heavy bill at a moment when, in this telling, the balance of power has shifted and the case for stronger action is sharper than before. The author argues that the company seems comfortable with limits that would now be unacceptable, and that Cimino’s reply never really engages that point.

The dispute centers on four areas: preemption, model blocking, risk thresholds, and updates. On preemption, Anthropic says state law should only be overridden if Congress sets a very strong federal standard, while still preserving room for states on issues like children’s protections, whistleblower laws, and liability. The reply says the language will be revised to make state powers clearer. The critique is blunt: whatever Anthropic says it meant, the draft still does not read like a strong federal standard.

Model blocking is the sharpest split. Anthropic says Commerce should be able to go to federal court to stop a dangerous model, backed by an independent evaluator, but wants a meaningful legal threshold so the power is not abused. The response says courts should be able to move quickly in emergencies. The criticism is that this is still too cautious for frontier AI, where the damage can be irreversible and where a formal “take it down first, argue later” process would be better than the current ad hoc version.

On risk thresholds, Anthropic points to California law as the source of its draft standard and says those thresholds were always a compromise. But the author calls them too high, especially now that there are real-world examples showing how little triggers action. Anthropic also says it has updated the proposal after recent incidents, including letting a court halt internal use of a dangerous model and requiring reporting when a model escapes containment. Even so, the overall judgment here is that the new draft is still weaker, more cautious, and more aligned with Anthropic’s business interests than with the urgency of the problem.