Anthropic is quietly pitching its own AI safety law: when the regulated party writes the rules, the fine print matters

By Politico · October 16, 2026.

Anthropic representatives have quietly presented lawmakers on Capitol Hill with a draft bill to regulate frontier AI safety, according to Politico, which cites three people familiar with the effort and obtained the confidential text. The draft had not previously been made public.

A company spokesperson confirmed it to Politico: the bill incorporates the proposals of its "Advanced AI Framework," which Anthropic has been promoting to federal and state lawmakers since June. The spokesperson described it as "sensible policies" that Congress should pass.

According to the article, 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. That risk must stem from biological weapons, from the model evading its developer's controls, or from a cyberattack. Politico notes that the language mirrors that of California laws such as SB 53, signed last year. Politico's description of the draft, as it appears in the material, also indicates that it would limit the circumstances for blocking a model and would override some state AI safety laws (what is known in the U.S. as "preemption": federal law displaces state law).

Our take: the substance may be defensible, but the process and the design deserve scrutiny. First, the good. Concrete, measurable thresholds (damage figures, narrowly defined risk categories) are preferable to panic-driven regulation, and they fit our position in favor of evidence-based governance. A regime triggered by real catastrophic risks, and not by any political discomfort, also protects open innovation and research.

Now, the fine print. A threshold of "more than $1 billion" or "more than 50 deaths," combined with the requirement that the risk be "imminent," sets a very high bar. In practice, the state could only step in when the damage is all but done. It is the difference between an emergency brake and preventive oversight; the draft, as Politico describes it, appears to opt for the former. And preemption is, in general, the most sensitive point: replacing a patchwork of state laws with a single federal rule simplifies compliance, but if the federal floor is low, it weakens protections in the most demanding states.

Who is involved also matters. An industry leader drafting the rules for its own industry is neither new nor illegitimate: regulators need technical expertise. But readers should distinguish between a framework that benefits the public and one that also sets a comfortable barrier for those already in compliance. That it was pushed quietly, rather than in public debate, does not help build trust, even though Anthropic maintains it has been advocating these ideas for months.

Winners and losers: large companies with compliance capacity and a single national standard come out ahead; states that want to go further, and to some extent preventive oversight, may lose out. It should be noted that the article, in the available material, does not detail other sections of the draft or lawmakers' reactions, so it is worth waiting for the full text.

In the long run, AI can help us eradicate diseases and generate abundance, and getting there requires institutions that manage the transition well. A clear federal law based on real risks is a useful piece. The question is not whether it should exist, but whether the bar for intervention is the right one and who sets it. That debate should be public, not a conversation behind closed doors.