Anthropic Pushes States to Tighten AI Rules as Safety Debates Intensify

Anthropic is urging US states to move faster and adopt tougher artificial intelligence rules, arguing that transparency-focused measures passed in California and New York last year are no longer enough to manage the risks posed by the most powerful AI systems. The push matters because state lawmakers are now filling the federal vacuum on AI oversight just as frontier-model developers race ahead.

The company, which once helped champion some of the first state-level safety laws, is now backing stricter requirements in places such as Illinois and Massachusetts, including third-party audits and stronger enforcement powers for attorneys general. Anthropic says the pace of AI progress has outstripped the policy response, and that governments need stronger guardrails before the technology becomes even harder to control.

That position places the nearly trillion-dollar startup in an unusual role: both a leading commercial builder of advanced AI and one of its most aggressive advocates for regulation. The stance has earned praise from AI safety advocates, but it has also fueled accusations from rivals and critics who say Anthropic is using regulation to entrench its own market position.

Why Anthropic is pressing states for faster AI regulation

Anthropic says the main reason for its policy shift is simple: AI systems are getting more capable faster than lawmakers can react. The company argues that the early wave of safety laws was valuable, but insufficient for the current generation of frontier models.

Cesar Fernandez, Anthropic’s head of US state and local government relations, said the company believes the policy environment must now match the speed of technical change. In his view, basic transparency and self-reporting are not enough when the systems in question can affect critical infrastructure, public safety, and potentially even national security.

“The transparency-focused safety bills of 2025 were a really important start, but as the capabilities of AI systems continue to advance quickly—the policy responses need to match,” Fernandez said.

He added that, for the most powerful systems, transparency alone should no longer be treated as a sufficient safeguard. Anthropic wants states to go beyond asking companies to disclose what they are building and instead impose verification, audit, and enforcement mechanisms.

The company’s message also reflects a broader political reality: Congress has made little progress on AI legislation, leaving states to lead on a fragmented patchwork of rules. That has opened the door for companies, advocacy groups, and unions to lobby state legislatures directly, turning statehouses into a major arena for AI governance.

What laws has Anthropic supported so far?

Anthropic has backed several of the toughest state proposals aimed at frontier AI developers. The company supported California and New York transparency laws passed last year and has since expanded its endorsement list to include bills in Illinois and Massachusetts.

These measures are not identical, but they share a common goal: forcing large AI developers to document safety practices, disclose model risks, and submit to outside scrutiny. Anthropic says those rules should apply only to the biggest model builders, not to the broader population of smaller startups and app developers.

The shift toward stronger rules is notable because much of Silicon Valley opposed the earlier bills, warning that aggressive regulation could slow innovation. Anthropic took the opposite path, aligning itself with lawmakers and safety advocates who argue that frontier AI needs oversight before it is widely deployed.

How far does Anthropic want states to go?

Anthropic wants states to do more than require disclosures, but less than grant themselves sweeping authority to shut down AI systems. The company has drawn a line around who should have the power to block a model from being deployed.

In a policy document released last month, Anthropic said governments should have a mechanism to stop the release of unsafe AI models. But the company also argued that such a power should sit with the federal government rather than state legislatures, even as it supports stronger state-based oversight in other areas.

That position reflects a balancing act. Anthropic wants meaningful external checks on frontier models, yet it appears wary of a system in which 50 states can independently halt model launches. The company has suggested that if a model is so risky that deployment must be blocked, the process should be standardized, transparent, and nationally coordinated.

The issue is not abstract. The Trump administration recently ordered Anthropic to suspend access to two of its most advanced models for foreign nationals, forcing the company to take them offline for all users. Anthropic publicly criticized the move and later said any blocking of model deployment should occur through a fair evaluation process rather than ad hoc government action.

Why third-party audits matter

Third-party audits are central to Anthropic’s preferred state approach because they move beyond self-policing. The company argues that outside reviewers can provide a more credible assessment of whether a lab’s internal safety practices are actually robust enough for high-stakes systems.

Under that logic, the role of the state is not to design AI models but to verify that the companies building them are following standards that reflect their scale and risk profile. For Anthropic, that means compliance should be tied to the power of the model rather than the broader category of “AI company.”

What critics say Anthropic is really doing

Not everyone sees Anthropic’s campaign as a pure safety effort. Some critics, including high-profile Silicon Valley figures, say the company is using state regulation to make life harder for smaller rivals.

David Sacks, a former White House AI adviser and current technology adviser to President Donald Trump, has argued that Anthropic is trying to engineer a regulatory environment that advantages incumbents. In his view, burdensome rules would slow the broader startup ecosystem while leaving large, well-funded players better positioned to comply.

“Anthropic is running a sophisticated regulatory capture strategy based on fear-mongering,” Sacks wrote in a social media post last year.

He also accused the company of driving what he described as a damaging regulatory wave that harms innovation. That critique has found an audience among entrepreneurs and investors who worry that frontier-AI rules could function as a barrier to entry.

Anthropic rejects that framing. Fernandez said the company has backed only legislation aimed at “large AI model developers,” a category that is usually defined by very high spending thresholds and substantial annual revenue. He argued that it is difficult to imagine a small startup meeting those standards.

Even so, the line is getting blurrier as the cost of training frontier models rises. Several heavily funded companies, including Safe Superintelligence, Thinking Machines Lab, and Mistral, have raised billions of dollars and could eventually grow into the kind of firms these laws target. That possibility gives the policy debate an added competitive edge: regulations aimed at the biggest labs today may shape who can compete tomorrow.

How are frontier AI companies defined by these bills?

They are generally defined by scale, not by whether they are startups in the traditional sense. The laws Anthropic supports tend to focus on companies that have spent hundreds of millions of dollars on AI development and generate more than $500 million in annual revenue.

That definition is meant to capture the very largest model developers while sparing smaller firms and product builders from the most burdensome requirements. In practice, though, it could still sweep in fast-growing companies as soon as they cross the threshold.

• Spending on AI development is a key trigger.
• Annual revenue thresholds are used to separate large labs from smaller firms.
• The rules are designed to apply to frontier model developers, not every AI company.
• Third-party audits are intended to verify internal safety processes.

Anthropic’s view is that if a company is building a system powerful enough to pose systemic risk, it should not matter whether that company is a household name or a newly public startup. The risks, it argues, come from capability, not brand.

Who is Cesar Fernandez and why does he matter?

Cesar Fernandez is the policy executive now leading Anthropic’s state and local government strategy, and he brings years of experience in regulatory fights. Before joining Anthropic earlier this year, he worked in government-relations roles at FanDuel and Uber, helping both companies navigate state-level battles.

His background matters because the AI regulation debate is increasingly being decided in state capitols, where experienced lobbyists and policy specialists can shape bill language, enforcement mechanisms, and exemptions. With federal AI legislation stalled, companies are treating state advocacy as a core business function rather than a side issue.

Fernandez’s remit suggests Anthropic is making a long-term bet that state policy will remain central to AI governance. That is especially true in a period when lawmakers are still trying to understand the real-world risks of frontier systems and how to regulate them without freezing innovation.

Why federal inaction is pushing the fight to the states

The federal government has not produced a comprehensive AI regulatory framework, and that vacuum is one reason states have become so important. As Congress stalls, governors, attorneys general, and state lawmakers are increasingly setting the terms for what AI companies can and cannot do.

For Anthropic, that creates both an opportunity and a risk. The opportunity is to help build a rules-based environment that rewards safety and responsibility. The risk is that a patchwork of state laws could become inconsistent, costly, and politically volatile.

That tension explains why Anthropic is willing to support some state action while resisting others. It wants a strong enough framework to raise the baseline for frontier AI, but not so much fragmentation that deployment becomes unworkable.

How the policy debate has evolved

In the earliest phase of the state AI fight, the central question was whether lawmakers should require transparency at all. Now the debate has moved toward verification, auditing, and the possibility of intervention before deployment.

That evolution mirrors the broader development of AI itself. As models become more powerful, policymakers are shifting from asking what a company says it did to asking whether an independent party can confirm that safety claims are real.

What risks are lawmakers trying to prevent?

Lawmakers and safety advocates are focused on catastrophic scenarios that could stem from frontier AI systems, including financial disruption, infrastructure failures, and other forms of large-scale harm. Anthropic’s support for strict laws is rooted in the idea that once systems reach a certain capability threshold, the potential downside is too great to rely on voluntary compliance alone.

The company has repeatedly emphasized the possibility that advanced AI could contribute to severe societal damage if left unchecked. That framing has helped it win allies among groups that believe the public is underprotected against powerful AI labs.

At the same time, more familiar public concerns—such as job losses, data center impacts, and the effect of chatbots on children—have generated far less coordinated lobbying from frontier labs. Anthropic says it is open to broader discussions, but its most visible political effort remains centered on existential-risk prevention.

Fernandez said Anthropic is “eager to engage with lawmakers” on issues beyond catastrophic safety and is already discussing those topics across the states.

How does Anthropic’s position affect the wider AI industry?

Anthropic’s approach matters because it helps legitimize the idea that frontier AI should face heavier oversight than ordinary software. When a major lab supports state regulation, it gives lawmakers political cover and can reshape the terms of the debate.

It also creates pressure on competitors. Other AI companies may be forced to respond to the same bills Anthropic supports, even if they would prefer lighter-touch rules. In that sense, Anthropic’s advocacy is not just about compliance; it is also about defining the operating environment for the next generation of AI firms.

The company’s recent model releases have already put cybersecurity concerns in the spotlight, further strengthening the case that advanced AI should not be treated as a normal consumer technology. Anthropic has spent years warning legislators about the risks of frontier systems, and that sustained effort has given it unusual influence in state policy circles.

What happens next?

The most likely next phase is a continued state-by-state fight over who qualifies as a frontier model developer and what obligations those companies should face. As more states consider audit requirements, disclosure rules, and enforcement mechanisms, Anthropic will likely remain one of the most influential voices in the room.

The bigger unresolved question is whether the federal government will eventually step in with a national standard. If it does, some of the current state-level debate could be pre-empted or consolidated. If it does not, the patchwork approach will deepen—and companies like Anthropic will keep trying to shape it bill by bill.

For now, Anthropic’s position reveals a central paradox in the AI era: the companies building the most advanced systems are also among the loudest voices calling for rules to constrain them. Whether that is a sincere safety mission, a smart competitive strategy, or some mixture of both, it is already shaping the future of AI policy in the United States.

And with state lawmakers moving faster than Congress, the next major AI regulatory breakthroughs are likely to come not from Washington, but from the statehouses Anthropic is now actively courting.