Anthropic Pushes States for Tougher AI Safety Rules

Anthropic, the AI safety-focused startup now valued at nearly $1 trillion（約160兆円）, is taking an unusual step for a company of its stature by actively pushing U.S. states to adopt stricter AI regulations. The company's head of U.S. state and local government relations, Cesar Fernandez, told WIRED that the transparency-focused safety bills passed in 2025 were "a really important start," but that "as the capabilities of AI systems continue to advance quickly—the policy responses need to match." Anthropic argues that transparency and self-reporting are no longer sufficient safety measures for the most powerful AI systems.

Anthropichas already backed transparency requirements in California and New York last year, earning praise from AI safety groups and labor unions but drawing criticism from Silicon Valley leaders who view the effort as a competitive strategy. David Sacks, the former White House AI czar and current technology adviser to President Trump, accused Anthropic of running "a sophisticated regulatory capture strategy based on fear-mongering" to trap smaller startups in red tape. More recently, Anthropic supported an Illinois measure requiring AI labs to undergo safety evaluations by third-party auditors, and endorsed a Massachusetts policy requiring similar audits while empowering the state attorney general to seek injunctive relief against noncompliant companies.

Fernandez rejected the regulatory capture accusation, noting that Anthropic has only supported bills targeting "large AI model developers"—a term that varies by statute but generally refers to companies that have spent hundreds of millions of dollars on AI development and have more than $500 million（約800億円） in annual revenue. "It's hard to imagine a startup meeting that threshold," Fernandez says. However, in today's capital-intensive AI landscape, several well-funded competitors—Safe Superintelligence, Thinking Machines Lab, and Mistral—have each raised billions of dollars from investors, though their revenue remains far lower than Anthropic or OpenAI. Fernandez contends that the more defensible argument is that any company large enough to develop a powerful AI model should face the same regulations, because the underlying risks are the same. He frames Anthropic's goal as inspiring "a race to the top developing the most safe and secure AI systems."

Anthropic has been clear about where it will not accept regulation. In a policy document published last month, the company recommended that governments should have a mechanism to block unsafe AI model deployments—but only at the federal level, not state level. This position echoes Anthropic's resistance to the Trump administration's recent directive, which led the company to suspend access to its Mythos and Fable 5 models for foreign nationals. Anthropic argued in a blog post that blocking AI model deployments should occur only through a fair, transparent evaluation process. Fernandez noted the position remains an evolving conversation. Beyond state-level work, Anthropic has also sent a letter to the U.S. government accusing Alibaba of a "distillation attack"—extracting information from Anthropic's models through systematic prompting to develop competing tools. Some AI researchers have dismissed the claim as regulatory capture by another means, arguing Anthropic aims to persuade the government to ban open-source Chinese models, which could push American businesses toward Anthropic's enterprise offerings. Notably, Anthropic CEO Dario Amodei has previously warned Congress about the dangers of open-source AI. Fernandez clarified that Anthropic has not specifically targeted open-source models in endorsed state legislation, saying the focus is on model capabilities and whether they rise to a level requiring regulation.

The gap between Anthropic's aggressive push on catastrophic AI safety rules and its more muted efforts on voter concerns is striking. Voters are more frequently concerned about job losses from AI, negative impacts of data centers in their communities, and effects of chatbots on children—issues that have not inspired comparable legislative campaigns from frontier AI labs. Anthropic and rivals have made promises that ordinary taxpayers won't bear data center costs, and Anthropic has published proposals for responding to AI-driven job displacement, but the industry has applied considerably less political muscle to state laws addressing those problems. When asked about the discrepancy, Fernandez said Anthropic is "eager to engage with lawmakers" on issues beyond catastrophic risks and is already "having those conversations throughout the states," though those discussions have not yet produced the kind of coordinated state-level push the company has mounted around existential AI risks.

Anthropic's embrace of stringent AI regulation sits at odds with its position as a high-valuation startup. The company was founded with the mission to ensure the world safely navigates transformative AI, and as it has grown into a major player, it has leaned into supporting some of the nation's harshest proposed rules on frontier AI companies. These rules are framed as mitigating catastrophic risks—including the possibility that advanced models could contribute to financial disasters or mass deaths—though critics in Silicon Valley interpret the political strategy differently.

The tension centers on the definition of "large AI model developer." Anthropic's Cesar Fernandez argues that the thresholds in supported legislation (generally companies with hundreds of millions in development spending and over $500 million（約800億円） in revenue) are so high that startups cannot reach them easily. Yet in a capital-intensive AI landscape, a handful of well-funded competitors—Safe Superintelligence, Thinking Machines Lab, and Mistral—have already raised billions of dollars. The accusation from figures like former White House AI czar David Sacks is that Anthropic is using regulatory frameworks to entrench its own advantage while appearing to act in the public interest. Anthropic counters that any company powerful enough to build dangerous AI systems should face the same regulatory burden, framing it as a "race to the top" in safety rather than a competitive move.

Where Anthropic draws firm lines is revealing. The company opposes giving states (rather than the federal government) the authority to block AI model deployments, and it resisted the Trump administration's export controls on its own Mythos and Fable 5 models. This suggests that while Anthropic is willing to accept robust auditing and transparency requirements, it wants to avoid fragmented, state-by-state deployment restrictions that could impede its business. Meanwhile, the company has mounted far less political pressure on regulations addressing voter concerns like job displacement or data center impacts—issues that would require greater structural compromise from the entire industry.