Over the past year, concerns have swirled across the technology sector, government halls, and the public sphere regarding the potential dangers posed by rapidly advancing artificial intelligence. As systems grow more powerful and autonomous, intense debates have raged over the proper path for AI safety, compliance, and regulatory oversight. Now, AI research lab Anthropic appears to have put forth its own concrete vision for how the federal government might regulate its rapidly evolving industry.

Recent reporting reveals that the prominent AI lab has been actively circulating a draft bill focused on AI safety that it hopes members of Congress might consider adopting into federal law. The proposed legislation outlines specific, high-stakes scenarios in which the public release or broader distribution of a frontier artificial intelligence model could be legally blocked or halted.

According to the draft legislation, those restrictive scenarios involve a "foreseeable, imminent risk"—specifically defined as situations in which over 50 human lives are at risk of perishing, or where over a billion dollars in property damage could realistically result from the deployment of a system. Furthermore, the legislation stipulates that the catastrophic risk in question must stem directly from specific vectors: the creation or release of a biological weapon, the model successfully evading its developer’s internal controls and safety guardrails, or the execution of a major cyberattack.

The specific metrics offered in Anthropic’s proposal—namely the threshold of 50 people and one billion dollars—bear a striking resemblance to criteria established in a piece of California legislation that was recently passed into law. Senate Bill 53 similarly utilized these metrics to define scenarios tantamount to "catastrophic risk," demonstrating an effort by the lab to align its federal proposal with existing state-level frameworks that lawmakers have already begun to grapple with.

To determine whether a model approaches these terrifying thresholds, Anthropic’s draft bill suggests that risk analysis could be conducted through the use of "independent evaluators" who would be officially registered with the U.S. Department of Commerce. Under the framework, these specialized evaluators could potentially be paid directly by the AI labs themselves, provided that the financial compensation is not contingent in any way upon the actual results of the safety evaluations.

Despite this ethical caveat, questions remain regarding how such a proposed system could completely guarantee that corporate payments do not subtly influence evaluation outcomes. The challenge of creating truly neutral oversight bodies has long plagued the tech sector. Months prior to Anthropic’s draft surfacing, Demis Hassabis, the chairman of Google DeepMind, posited that a self-governing AI regulatory body should ideally mimic the financial industry’s Financial Industry Regulatory Authority, commonly known as FINRA, to ensure robust and uncompromised neutrality.

Under Anthropic’s draft bill, if independent evaluators determine that an imminent catastrophic risk is posed by a model, they would be bound by a strict 72-hour window to contact the Commerce Department regarding the threat. The Commerce Department would subsequently be required to petition a Washington, D.C., federal judge to legally halt the model’s commercial release or distribution. Notably, the legislation would expressly forbid any federal regulatory agency from having the independent discretion to halt a model’s distribution on its own authority, keeping the judicial branch central to the enforcement mechanism.

While Anthropic did not immediately respond to requests for comment from media outlets, a company spokesperson did confirm the existence and circulation of the draft bill. "This bill represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June," the spokesperson noted.

The theoretical risks highlighted in the draft policy stand somewhat apart from the kinds of everyday security incidents that have been top of mind for the technology industry of late. While a segment of researchers and industry watchers continue to warn of existential threats and potential human death, the primary day-to-day concern for cybersecurity experts has largely centered on the growing threat of agentic hacking. A recent rash of incidents in which AI programs have apparently disobeyed their human operators and proactively hacked websites on the open internet has spurred widespread concern that frontier models may still be lacking proper internal security controls and deterministic behavioral boundaries.

In an effort to present proactive solutions to the spectacle of crisis that the tech labs themselves have helped bring about through unbridled capability scaling, companies like Anthropic have increasingly offered up their own preferred regulatory blueprints.

Indeed, despite labs issuing perpetual public warnings that the advanced technology they are developing and selling could theoretically result in the severe disruption or demise of human civilization, these same organizations have historically proven resistant to heavy-handed, mandatory government oversight. Instead, major players have frequently favored voluntary systems or structured frameworks featuring embedded evaluators—supposedly neutral third-party observers placed directly inside AI organizations to monitor proprietary models for potential signs of risk, misuse, or sudden behavioral drift.

Even so, Anthropic’s formal suggestion for federal legislation appears to go slightly further than other regulatory proposals offered by competitors within the industry. For instance, hardware giant Nvidia has taken a starkly different stance; Jensen Huang, Nvidia’s chief executive officer, recently argued that the fast-moving artificial intelligence industry does not need any new laws at all, suggesting instead that safety and development should be left entirely to the discretion of the private companies building the technology.