Anthropic's draft bill sets the intervention bar at 50 deaths or $1B in damage: safety or regulatory shield?

By Mother Jones · October 17, 2026.

Anthropic has begun lobbying the U.S. Congress for a draft bill to regulate advanced AI, according to Politico, as reported by Mother Jones. The text would allow the government to block the release or use of a model only if it poses a "foreseeable and imminent" risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people.

The scope is limited to three risk categories: the creation of biological weapons, a model evading its developer's controls, and cyberattacks. Other harms, such as AI-enabled fraud, would be excluded, according to Mother Jones' description. The draft would also reportedly bar any agency other than the Department of Commerce from halting a model's development or internal use.

An Anthropic spokesperson told Politico that the text reflects its "Advanced AI Framework," which the company has been promoting to federal and state lawmakers since June, and described it as "sensible policies" that Congress should pass.

The backdrop is mounting pressure. Mother Jones notes that lawmakers from both parties have proposed heavily taxing AI companies, pausing the development of frontier models or creating a cabinet-level AI oversight department. According to the outlet, Anthropic spent more on lobbying Congress in the first half of 2026 than in all of 2025. Mother Jones adds that the measure could prevent states from passing stricter laws, although that is the outlet's interpretation, not text we have been able to review.

A note on the source: Mother Jones has an editorial line critical of the industry, and its piece combines Politico's reporting with its own assessment. We have not seen the full draft, so the details should be treated as "as reported."

**Our take.** The criticism has merit, but the underlying question is a different one. A threshold of 50 victims or $1 billion is not absurd if the goal is an "emergency brake" for catastrophes: blocking a model is a drastic measure, and a law allowing it in the face of any minor risk would be unenforceable and would likely risk regulating panic rather than actual capability. That is a position we support: evidence-based governance.

The problem is that an emergency brake is not an oversight regime. If the only power is a veto when harm is imminent and massive, everything that comes before is left uncovered: pre-release testing, transparency, audits, minor incidents. And, as we have argued, the most immediate danger of agentic AI tends to be fraud and the automation of abuse at scale, precisely what the draft, according to the reporting, leaves out. Limiting risk to three categories is consistent with the most severe scenarios, but it ignores cumulative, everyday harms.

There is also an obvious incentive. Concentrating authority in a single agency and potentially preempting tougher state rules is a form of regulation that reduces uncertainty for whoever proposes it. The fact that Anthropic is a company that talks insistently about safety does not exempt it from that logic: an industry that sets its own bar has an informational advantage and a conflict of interest, and the bar deserves independent scrutiny, whatever the company.

Total cynicism isn't warranted either. A clear federal framework, with a single point of contact and explicit criteria, may be better than a patchwork of 50 contradictory state rules, and putting concrete thresholds in writing is a concrete starting point for negotiation. The sensible approach is to treat this draft as an opening offer, not the final law.

In the long run, AI's promise —curing diseases, generating abundance— depends on society trusting those who govern it. That trust is built through oversight that covers both extreme and everyday risk, and by keeping the watchdog separate from the developer. The test will be whether Congress expands the draft or adopts it as is.