Anthropic’s Proposed AI Safety Bill Only Kicks In at 50 Deaths or $1 Billion in Damages - Gadget Review
Fifty deaths. One billion dollars in damages. These are the numbers Anthropic believes should lead the federal government to step in when a dangerous frontier AI model poses a serious threat.
Anthropic has privately shared a draft federal AI safety bill with U.S. lawmakers, according to Politico's October 16, 2026, report. One of the draft's most important features is how it assigns power: the Commerce Department would be the only agency allowed to enforce the rules and seek to stop a model. The reported wording would prevent every other federal agency from independently suspending a model.
What It Takes to Stop a Model
The proposal sets a clear threshold for extreme harm that must be met before the government can step in against a frontier AI model.
According to Politico, the government could step in only if a frontier AI model poses a "foreseeable and imminent risk" of causing more than $1 billion in damage or killing or seriously injuring more than 50 people.
Three specific situations would meet the requirements: creating or releasing a biological weapon, carrying out an cyberattack powered by AI, or a model escaping the controls set by its developer.
The list of qualifying situations is deliberately limited. Under the proposal, there would be no clear way to suspend a model for a harmful AI incident that causes less damage than the stated limits or does not fit one of those three categories.
One Agency Controls the Process
Only the Commerce Department could ask to stop a model, and a federal judge would have to approve the request.
Under the proposed system, independent third-party evaluators would register with the Commerce Department and regularly check frontier AI companies and their models, according to Politico. AI laboratories could pay these evaluators directly, as long as the payments did not affect their independence. Whether this condition could be enforced in real life remains an open question about how the system is designed. The proposal does not establish that the condition would be enforceable.
If an evaluator found an immediate risk that matched one of the listed situations, it would have 72 hours after finishing its assessment to inform the Commerce Department, according to Politico. The department would then present the evidence to a federal judge at the U.S. District Court for the District of Columbia and ask for a court order to stop the model from being released or used.
The reported wording would prevent other federal departments, agencies, and officials from independently stopping the development, release, or internal use of an AI model. Giving one agency this power could make the process more predictable, but it would also put the responsibility in one place. If the Commerce Department chose not to act or could not assess a threat quickly enough, no other federal agency could seek to suspend the model under this bill. However, the available reporting does not show that every other type of federal action would be blocked. This is a possible result of how the enforcement system is designed, not a finding directly stated in the proposal.
Anthropic's Position
The company says the draft builds on its Advanced AI Framework, which it has been presenting to lawmakers since June.
An Anthropic spokesperson said the draft follows the company's Advanced AI Framework, which it has promoted to federal and state lawmakers since June. The company described the proposed measures as sensible steps that Congress should put into law.
According to Politico, the proposal would also require frontier AI laboratories to create safety plans, publish reports explaining their work and practices for new models, protect people who report wrongdoing, and tell the Commerce Department about serious safety incidents.
A Note on State Authority
The draft could limit some state-level AI rules in areas already covered by federal law.
According to Politico, the proposal would reportedly override some state AI laws that cover public reports about risks, third-party evaluations, and the reporting of serious safety incidents. This could be an important limit because some states are actively creating their own rules in these areas.
The proposal would have no legal force unless Congress considers it and passes it into law. Still, it clearly shows how one major AI company believes federal oversight should work: the government would act only when the risk reaches an extreme level, one agency would control enforcement, and a judge would have to review requests to stop a model. Court review would add oversight, although it would also take time. Whether these rules offer enough protection is likely to be a key issue in the debate among lawmakers in Washington.