Anthropic's leaked draft AI bill

Mowshowitz publishes critique of Scott Alexander's Against SURA, contesting arguments against unilateral AI development stoppage

The Washington Post reported that Anthropic's AEGIS Act draft departs from the FAA-style regulatory model Amodei proposed in summer 2026, and that the bill would let labs write their own compliance standards.

Kelsey Piper argued Anthropic is internally divided, with its policy team described as having gone rogue and hard structural constraints progressively replaced by reliance on leadership judgment. Nate Soares of MIRI agreed Anthropic is not meeting its stated burden on recursively self-improving AI research. Revised preemption language remained unpublished.

Miles Brundage highlighted a Semafor poll finding that voters in key states support AI safety measures.

Dean Ball wrote on October 28 that states now appear poised to pass meaningfully different AI laws imposing obligations on frontier AI developers, with federal inaction driving the divergence.

The piece adds weight to a central dispute over the AEGIS Act, a draft federal AI bill Anthropic has been circulating, whose preemption clause overrides state AI safety requirements in California, New York, and Illinois: critics have argued the bill strips states of safety authority while providing a weaker federal replacement.

After Amodei listed nine costly pro-safety decisions and accepted three points as communication failures in the AEGIS Act debate, critics held their positions: Calvin maintained Anthropic systematically prevents governments from constraining it, Kelsey Piper argued the sacrifices reflect Amodei's personal convictions rather than institutional governance, and commentators surfaced a 2022 Amodei statement that the company should not be racing ahead, contradicting his current claim.

The FTC separately demanded testimony from Anthropic executives.

TechTimes detailed AEGIS Act provisions allowing developer-selected auditors and preempting California, New York, and Illinois state laws.

On October 25, 1,324 employees of frontier AI companies asked the US government to support international tools to deliberately pace frontier AI development, stopping short of calling for an immediate pause.

Brendan McCord argued effective pacing would require illiberal enforcement mechanisms, while Samuel Hammond countered that models already breaking encryption and executing multi-step exploits make it a necessary response. Bill Gates separately proposed that open-weight models capable of sophisticated cyberattacks should run only on government-certified platforms, with developers free to use and customize them within certified environments.

The FTC issued civil investigative demands to Anthropic and OpenAI, adding regulatory enforcement to a fight that began when a bill Anthropic was circulating to Congress leaked and drew criticism for a high threshold on blocking dangerous models.

Reuters reported OpenAI president Greg Brockman and his wife donated $25 million to a group opposing AI regulation in more than 48 congressional races, while OpenAI separately asked California to strengthen its SB 53 AI safety bill. Meta withdrew roughly $25 million from its planned AI political spending.

Mowshowitz assessed Amodei's response as better than the initial one but said the controversy remained unresolved.

Anthropic CEO Dario Amodei publicly admitted the company's circulated draft federal AI safety bill was among three recent moves that were carelessly written or poorly communicated, accepting responsibility as CEO.

He listed nine commercially costly pro-safety decisions, including strict biological safeguards and restricting sales to PRC-controlled companies, and defended Anthropic's continued acceleration as core to a strategy of raising industry-wide safety standards by remaining at the frontier.

Anthropic's head of federal affairs Anthony Cimino issued a public defense framing the circulated draft bill as a translation of the company's existing safety framework, not a final proposal.

Mowshowitz rejected that framing, writing the bill is 'clearly sculpted to optimize for Anthropic business interests' and that 'Anthropic is not living up to its own ideals,' calling for its formal disavowal. Calvin argued Anthropic's actions, including RSP V3 and compute deals, are most consistent with prioritizing the AI race over safety.

After critics wrote that Anthropic's AI safety legislation would make it harder for courts to halt dangerous models than existing law allows, the company's head of federal affairs acknowledged the preemption language was 'not precise enough' and committed to revising it.

Democratic Senators Warner, Schatz, and Kim separately introduced the Artificial Intelligence Risk Management and Security Act of 2026 on September 29, which would establish a federal AI Safety Board. Nvidia CEO Jensen Huang also said he would not sell GPUs to AI labs he considers irresponsible.

Politico reporting corroborated that Anthropic's draft AI safety bill would confine halt orders to a single federal court in Washington D.C. and require evaluators to register with the Commerce Department.

Anthropic frames the bill as enabling oversight while mandating transparency. A UK parliamentary committee heard witnesses warn of a 50/50 chance AI could destroy humanity, but multiple MPs checked phones or left early, with the author contrasting this with American legislators described as more engaged on AI existential risk.

Safety advocates criticized on October 17 a leaked draft AI bill Anthropic is reportedly lobbying lawmakers to adopt, saying it is weaker than the FRONTIER Act on rulemaking and emergency shutdown.

The bill would block states from enjoining AI deployments even to prevent imminent catastrophic harm, setting the federal intervention bar at proof of imminent risk of killing more than 50 people or causing $1 billion in damage. David Sacks also accused Anthropic of using safety concerns to avoid antitrust scrutiny; Anthropic disputed that.

Anthropic's international chief Guillaume Princen argued at the International AI Summit in Dublin that open-weight models carry more safety risk by definition.

MilkRoadAI reports 45 AI models released in September 2026, described as the highest monthly total on record

Dean Ball argues law-following alone is insufficient as an AI alignment framework

Alignment Forum post proposes labs be prohibited from training a more capable model generation until current-generation alignment issues are resolved

Sacks and others compared Anthropic's lobbying to SBF's crypto regulatory strategy