Anthropic lobbies for AI safety bill that triggers only above $1B in damages or 50 casualties

What the proposal would do

Anthropic is quietly lobbying for federal AI safety legislation whose most powerful restriction would almost never apply. According to Tom's Hardware, citing reporting from Politico, the draft bill would allow the US government to block the release of an AI model only when that model poses a risk of causing at least $1 billion in damages, or of killing or injuring more than 50 people.

That is a deliberately catastrophic bar. Anything short of mass casualties or billion-dollar losses would fall outside the law's blocking mechanism, leaving the overwhelming majority of current and foreseeable model deployments untouched by its strongest provision.

Enforcement through the Commerce Department

The oversight mechanism described by Politico works as follows: an independent third party, registered with the Department of Commerce, would periodically evaluate the AI lab and its models. The lab under review could pay those evaluators, but only on the condition that the payment does not affect their independence.

Tom's Hardware highlights a structural consequence of this design: it would effectively make the Commerce Department the only federal agency capable of stopping an AI model. Whatever authority other regulators already hold, this framework would concentrate the specific power to halt a release in a single corner of the federal bureaucracy.

Anthropic's own framing

An Anthropic spokesperson told Politico, according to Tom's Hardware, that the bill reflects the proposals in the company's Advanced AI Framework, which the company has been advocating to federal and state lawmakers since June. The spokesperson described them as sensible policies that Congress should act on.

The notable point is that Anthropic is not resisting regulation here — it is actively campaigning for a specific version of it, one with thresholds and an enforcement architecture of its own design.

Why it matters

First, the numbers. If enacted, these thresholds would write into US law a definition of unacceptable AI risk that begins at mass-casualty and billion-dollar scale. Models capable of causing serious but lesser harm — biased decisions, fraud, privacy violations, targeted misuse — would sit below the line where any blocking power exists. The bill is aimed exclusively at catastrophic risk, not at the everyday failures that dominate current AI complaints.

Second, the concentration of authority. Making Commerce the sole agency able to stop a model means the practical question of whether a release can be halted would rest with one regulator and its registered evaluators, rather than being distributed across agencies with different mandates and appetites for intervention.

Third, the funding arrangement. Evaluators paid by the lab they assess, with independence preserved only by caveat, mirrors long-running debates in financial and environmental auditing about whether auditees should choose or compensate their auditors. Much would depend on how registration with Commerce polices that independence in practice.

Finally, there is the precedent. One of the largest AI labs is helping shape the safety rules it would then operate under. For model releases, the practical effect would be clarity: a lab that clears periodic third-party review, and stays below the harm thresholds, would have a strong argument that its release cannot legally be blocked anywhere in the federal system.