Critics Slam Leaked Anthropic AI Regulatory Bill as Weak and Dangerous

Overview

A draft of a proposed federal artificial intelligence regulatory bill, allegedly developed by Anthropic and circulated among lawmakers, has drawn sharp criticism from safety advocates and legal experts. The leaked document, which reportedly remains in a preliminary "bugs" phase, is viewed by many as insufficiently robust to address the growing risks posed by frontier AI systems.

Critics argue that the legislation prioritizes corporate interests over public safety, offering transparency measures without meaningful enforcement mechanisms. The timing of the leak has intensified concerns that the company is attempting to use a brief legislative window to preempt stronger state-level regulations.

Key Criticisms and Structural Concerns

The leaked bill shares similarities with the previously proposed FRONTIER Act, such as enhanced transparency provisions and stronger preemption of state laws. However, analysts note that it weakens critical enforcement powers, particularly regarding federal rulemaking and emergency shutdown authority.

"It reads as intentional, that Anthropic is more afraid that the government might enforce rules against Anthropic, than they are worried about the ability of the government to enforce rules against everyone."

Zvi Mowshowitz, a prominent critic of the industry's regulatory approach, described the bill as unacceptable. He argued that the proposal reflects a strategic effort to shield the company’s business interests while guarding against political risks, effectively placing corporate safety concerns above broader catastrophic risk mitigation.

Mowshowitz emphasized that there was no necessity for an organization purporting to prioritize AI safety to propose legislation that does not exceed existing compromise bills like the FRONTIER Act. He characterized the draft as resembling policy proposals typically pushed by anti-regulatory factions rather than safety advocates.

Preemption and State Powers

One of the most contentious aspects of the draft is its extensive preemption provisions. The bill appears designed not only to override specific state AI laws, such as California’s SB 53, but also to strip state courts of their general equitable powers to enjoin defendants and protect the public.

Critics highlight a particularly alarming clause that seems intended to prevent states from enforcing laws or judicial decrees that "suspend the development, deployment, or internal use of an AI model." Under this framework, the authority to block and suspend model deployments in the name of public safety would rest exclusively with the federal government.

This centralization of power has raised alarms among legal scholars who warn that it undermines the traditional role of states as laboratories for democracy and protectors of their citizens' welfare.

Federal Enforcement Standards

The federal government’s ability to restrict frontier AI development under the proposed bill is subject to an exceptionally high legal threshold. Authorities would only be able to intervene in cases involving an imminent risk of catastrophic harm, defined narrowly as events causing more than 50 deaths or over a billion dollars in damages.

Nathan Calvin, another industry observer, expressed visceral anger at the combination of far-reaching preemption and weak federal enforcement standards. He argued that passing the bill in its current form would leave the public less protected against catastrophic risks than existing laws currently do.

Calvin noted that while there are some sensible components, the overall package fails to meet the urgency of the moment. The strict criteria for federal action effectively render the government powerless to act against emerging threats that do not yet meet the bar for mass casualties or massive financial loss.

Industry Reaction and Comparative Analysis

Responses from within the AI policy community have been mixed but predominantly skeptical. While some observers acknowledged that the drafting quality was competent, they cautioned against granting the company too much credit.

Charlie Bullock, for instance, described the bill as "pretty reasonable" only when held against the standards of other frontier labs that have proposed no legislation at all. However, he stressed that the bill regresses in key areas compared to the FRONTIER Act, which was introduced back in June.

Bullock pointed out that the main deficiencies lie in two areas:

• Rulemaking authority, which is very strong in the FRONTIER Act but almost nonexistent in Anthropic’s draft.
• Emergency shutdown authority, which critics argue is dialed back significantly.

He suggested that while dialing back these powers might make sense from a perspective of distrusting the executive branch, it is ill-advised given the current political momentum toward safety regulation.

Improvements and Limitations

Despite the widespread criticism, the draft does contain certain improvements over previous proposals. The bill specifies that auditors are employed on a continuing basis rather than on an ad-hoc model, which could ensure more consistent oversight. Additionally, it offers better provisions regarding auditor access to information, although some experts argue these access rights are still undercooked and raise potential Fourth Amendment concerns.

Nathan Calvin acknowledged these positive aspects but maintained that they do not offset the structural weaknesses of the bill. He expressed hope that future embedded auditor bills introduced in the coming months will be significantly more ambitious and better aligned with treating catastrophic risks as serious national security issues.

Implications for the Legislative Landscape

The leak has sparked urgent calls for Anthropic employees and the broader public to scrutinize the document closely. Mowshowitz, writing from a conference, emphasized the importance of disseminating this information quickly so that stakeholders can demand accountability or clarification from the company.

If the allegations are true, this incident marks a significant shift in how major AI developers approach regulation. Rather than supporting comprehensive federal frameworks, companies may be quietly lobbying for lighter-touch regulations that preempt stricter state laws while limiting federal enforcement capabilities.

The political timing adds another layer of complexity. With a Republican Congress and a perceived "lame duck" session, critics fear that pro-industry groups are attempting to fast-track legislation before stronger safety measures can gain traction.

"Say it isn't so, Anthropic, or at least that it was a mistake and is no longer so, or I am going to be rather cross."

As the debate intensifies, the tech industry faces increasing pressure to align its policy advocacy with its stated commitment to AI safety. Whether this leaked draft represents a genuine policy proposal or a negotiating position remains unclear, but its reception suggests that any final legislation must do considerably more than simply preempt state laws without offering robust federal protections.