Anthropic's State Ratchet: AI Safety as a Moe

Anthropic's State Ratchet: Why the AI Safety Leader Wants 50 Rulebooks, Not One

Anthropic is deliberately engineering a state-by-state regulatory ratchet — each bill stricter than the last — as a competitive weapon against OpenAI and a hedge against federal preemption. The strategy turns "reverse federalism" into a race to the top that raises rivals' costs faster than Anthropic's own.

Anthropic is pursuing a strategy of regulatory one-upmanship, pushing each successive state to pass tougher AI safety rules than the last — California's transparency law in September 2025, New York's penalty-backed RAISE Act, Illinois's first-of-its-kind third-party audit mandate signed this month, and now a Massachusetts proposal the company calls the nation's strongest. The approach, laid out by Cesar Fernandez, Anthropic's head of U.S. state and local government relations, in a July 15 interview with POLITICO reported by Business Insider, is a direct rebuke of rival OpenAI's push for a single harmonized state framework — and a bet that a rising floor of state obligations will lock in the compliance moats Anthropic has already built while hobbling the Trump administration's attempt to freeze state action from Washington.

Two labs, two theories of state power

The split is doctrinal. OpenAI's top lobbyist, Chris Lehane, has coined "reverse federalism" to describe his campaign to bypass a paralyzed Congress by mirroring one model bill — California's — across statehouses, building a de facto national framework from the bottom up. OpenAI spokesperson Liz Bourgeois defended the approach to Business Insider, arguing it "helps regulators enforce the law, gives the public clearer protections, and allows developers to focus resources on safety rather than conflicting requirements."

Anthropic's Fernandez wants the opposite trajectory. "We're looking for legislation that meaningfully raises the bar on safety for the most capable AI systems," he told POLITICO. "Transparency and self-reporting, we don't believe are sufficient anymore." Each bill Anthropic backs is designed to be stronger than its predecessor, ratcheting obligations upward rather than converging around a single ceiling. The veiled jab is on-brand: Anthropic's founders left OpenAI in 2020 over safety disagreements, and the company was the only leading lab to endorse California's SB 53 before its passage, according to a letter organized by Rep. Doris Matsui and signed by dozens of House members opposing federal preemption efforts Matsui House letter.

The stakes are structural. With Congress deadlocked and the White House oscillating between a light touch and a heavy hand, the shape of U.S. AI rules is being written in Sacramento, Albany, Springfield, and Boston — and the two labs that dominate frontier model development are now spending political capital to steer those statehouses in opposite directions.

The California precedent — and its limits

California's SB 53, signed September 29, 2025, requires developers of frontier models, defined as those trained above 10²⁶ floating-point operations with developers earning over $500 million in annual revenue, to publicly report their safety governance practices. According to Brookings, approximately five to eight companies fall under its jurisdiction, including OpenAI, Anthropic, Google DeepMind, Meta, and Microsoft. The law includes a novel federal-deference clause: if a company satisfies comparable federal standards, like those in the EU AI Act, California will accept that compliance instead of requiring duplicate filings.

But Brookings analysts describe SB 53 as "light touch regulation" — a "structured transparency" regime designed to build an evidence base rather than impose hard limits. Laura Caroli, senior fellow at CSIS's Wadhwani AI Center, called it "modest but reasonable" and found its reporting requirements similar to voluntary agreements tech companies had already signed at the Seoul AI summit, as Al Jazeera reported. The law creates no mechanism for blocking the release of a dangerous model. It requires disclosure, not restraint.

This is the gap Anthropic's ratchet is designed to close. New York's RAISE Act, passed by the state legislature in June 2025 and still in negotiation with Gov. Kathy Hochul, goes further: it requires frontier developers to publish safety plans, disclose critical safety incidents to state authorities, and refrain from releasing models that fail their own internal safety tests. Penalties reach $10 million for a first violation and $30 million for subsequent ones — a figure that makes the California regime look like a disclosure exercise by comparison. New York Assembly member Alex Bores, the bill's sponsor, told NPR the RAISE Act was "largely based on voluntary commitments that all the companies had already made," codifying obligations labs had already pledged to meet.

Illinois, signed into law this month by Gov. JB Pritzker, adds a new obligation: annual independent third-party audits of safety plans, a first-of-its-kind mandate. OpenAI joined Anthropic in backing the Illinois measure, even as it lobbied Hochul to soften New York's bill toward the California model, suggesting the industry's two giants converge when the marginal cost of compliance is manageable but diverge when the obligation bites.

Massachusetts is where Anthropic draws its sharpest line. In late June 2026, it endorsed language in an economic-development bond bill that the company calls the nation's strongest state AI safety proposal. The provisions: independent evaluators must assess the potential for catastrophic risks, including AI assisting in bioweapons development, and the state attorney general is empowered to enforce the mandate. Fernandez frames this as a response to model capability, not lobbying tactics: the rapid development of increasingly powerful systems, including Anthropic's own Claude Mythos model, drove the company to weigh in earlier and harder. Mythos, in testing, was found capable of exploiting security flaws in every major computer operating system.

The Mythos catalyst and the case for mandatory audits

The cybersecurity panic around Mythos is the thread connecting Anthropic's lobbying posture to its regulatory bet. In June 2026, the Department of Commerce's Bureau of Industry and Security imposed export controls on Anthropic's Mythos 5 and Fable 5 models, requiring an export license for any foreign person, including Anthropic's own foreign-national employees, to access them. According to CSIS analysis, the move followed reports that Amazon researchers had identified jailbreaking methods that bypassed Fable 5's guardrails, potentially allowing users to identify and exploit cyber vulnerabilities. Anthropic said the order forced it to "abruptly disable Fable 5 and Mythos 5 for all our customers."

The episode exposed a gap that voluntary safety commitments and transparency disclosures were insufficient to close: a model with national-security-grade cyber-offense capability was deployed, and the federal government had to scramble to contain it after the fact. That is precisely the gap Anthropic's state strategy targets. Fernandez told POLITICO that California-style "transparency and self-reporting" is no longer adequate. The ratchet toward mandatory independent audits and catastrophic-risk assessments is, from Anthropic's vantage, a direct response to the Mythos precedent — a model the company itself built and tested, and that the federal government then had to restrict.

The Brookings Institution described the administration's posture as a series of "haphazard lurches" — from adamantly opposing any regulation, to a 30-day voluntary review period, to effectively shutting down the most advanced models — and called on Congress to "enact meaningful AI regulation" rather than relying on executive improvisation.

The competitive logic: raising rivals' costs

This is where the strategic logic surfaces. Anthropic has already invested heavily in internal safety infrastructure, including red-teaming, responsible scaling commitments, and frontier model evaluations, practices that predate any legal mandate. A regulatory regime requiring independent third-party audits and catastrophic-risk assessments imposes costs on all frontier developers, but those costs fall disproportionately on labs that have not built comparable internal capacity. Anthropic, having done the upfront work, is positioned to absorb the compliance burden more cheaply than competitors racing to catch up.

OpenAI's "reverse federalism," pushing a single harmonized model bill that mirrors California's lighter-touch transparency regime, looks in this light like a strategy to cap the obligation ceiling at a level OpenAI can meet without restructuring its safety practices. Anthropic's ratchet is designed to push that ceiling higher, state by state, in a way that raises rivals' costs faster than its own.

The historical parallel is the California effect in environmental regulation. Under the Clean Air Act, California won a waiver to set auto emissions standards stricter than federal minimums. Because automakers could not economically build two fleets — one for California and one for the rest of the country — California's standards became the de facto national floor. Thirteen states representing over a third of the U.S. auto market adopted California's rules, forcing manufacturers to meet the strictest standard everywhere. The AI equivalent is even more binding: you cannot build two versions of a frontier model for different states. A frontier model is a single artifact, deployed everywhere at once. Strictest-state requirements become the national standard by default, and Anthropic is engineering that strictest state to be progressively stricter with each legislative session.

Washington's preemption problem

Both strategies operate in the shadow of federal preemption — or the lack of it. Congress has passed no comprehensive AI legislation. A Congressional Research Service report noted that as of late April 2025, at least 48 states and Puerto Rico had introduced more than 1,000 bills containing the term "AI" in the 2025 legislative season, according to CRS Report R48555. The Trump administration has tried twice to freeze state action and failed both times. A 10-year moratorium on state AI laws was stripped from the One Big Beautiful Bill Act by a 99–1 Senate vote. A subsequent attempt to insert preemption language into the 2026 National Defense Authorization Act also failed.

On December 11, 2025, Trump signed an executive order creating an AI Litigation Task Force to challenge "onerous" state laws and directing Commerce to study withholding federal broadband funding from noncompliant states, as NPR reported. White House AI adviser David Sacks said the administration would not oppose state laws on children's safety but would push back on the "most onerous" examples.

The order is legally fragile. Legal scholars and state officials — including Republican governors Spencer Cox of Utah and Ron DeSantis of Florida — publicly questioned whether an executive order can preempt state legislative authority, a power the Constitution reserves to Congress. "An executive order doesn't/can't preempt state legislative action," DeSantis posted on X. John Bergmayer, legal director of Public Knowledge, told NPR the administration's theories "don't work very well" legally. CSIS concluded the order "puts the cart before the horse," marshaling federal resources to block state laws "well before a durable framework for regulating AI is in place."

This gridlock is what makes the state-level battle decisive. Brookings' analysis of state AI bills introduced from January 2023 through October 2025 found that younger, wealthier, Democratic-leaning states are leading on AI legislation, while older, poorer, and conservative states are far less active, according to Brookings. The states where Anthropic and OpenAI are competing — California, New York, Illinois, Massachusetts — are precisely the high-capacity, Democratic-leaning jurisdictions where ambitious regulatory bills survive. Brookings notes that states with Democratic governors see concentrated pushes on AI regulation, and that highly educated states introduce more technically demanding legislation, which generates more stakeholder contestation but also more ambitious frameworks. Approximately two-thirds of state AI bills in 2025 were introduced by Democrats, according to Brookings.

Winners and losers

The winners are identifiable. Anthropic wins on three fronts: its safety investments become a regulatory asset rather than a cost center, its rivals face higher compliance burdens, and the ratchet makes any future federal law more likely to lock in a high floor rather than preemption down to a low ceiling. State attorneys general win new enforcement authority and the staffing and budget that come with it. Independent audit firms — the companies that will conduct the third-party assessments Illinois now requires and Massachusetts proposes — gain a new market. Safety advocates who have pushed for binding obligations over voluntary commitments see their agenda advanced through state legislation that Congress will not pass.

The losers are equally concrete. OpenAI's "reverse federalism" strategy depends on states converging around California's model; Anthropic's ratchet breaks that convergence by making each successive bill a new baseline rather than a copy. Smaller frontier labs, the companies below the five-to-eight covered by SB 53's thresholds but approaching them, face rising compliance costs without Anthropic's head start on safety infrastructure. The Trump administration's preemption agenda loses another lever. Every state that passes a stronger law makes the executive order's litigation strategy look more like whack-a-mole. And venture capital opponents of regulation, who have framed Anthropic's safety posture as regulatory capture designed to "hamstring competitors," as Business Insider noted, now have a concrete strategy to point to rather than a vague accusation.

Diplomat View

Anthropic's state ratchet is not lobbying — it is regulatory architecture. The company is using state legislatures to build a rising floor of obligations that makes its own safety investments a competitive moat and makes federal preemption politically harder with every bill that passes. The decisive variable is whether the ratchet compounds fast enough to outrun any future congressional effort to preempt state law with a single national standard. If three more states pass audit or catastrophic-risk mandates by the end of 2026, the political cost of preemption rises sharply: Congress would be stripping not just California's transparency regime but Illinois's audit mandate and Massachusetts's enforcement teeth — a much harder sell to governors of both parties.

The forecast turns on three conditions. First, if the Trump administration's AI Litigation Task Force actually sues a state — likely New York or Massachusetts — and wins on interstate commerce grounds, the ratchet stalls. Second, if Congress passes a preemption bill with a high enough floor (mandatory audits, AG enforcement) that Anthropic can accept, the ratchet has served its purpose and the battle moves to the federal rulemaking. Third, if a model from a competitor causes a documented catastrophic harm — a cyberattack facilitated by a frontier model, a bioweapons-related output — the political momentum behind the ratchet becomes unstoppable and the question shifts from whether to regulate to how fast.

What to watch next:

• Massachusetts economic-development bond bill with AI safety language — legislative session timeline through fall 2026; watch for floor vote and amendment attempts.
• Trump AI Litigation Task Force first lawsuit against a state — expected target is New York's RAISE Act once Gov. Hochul signs a final version; filing would trigger Commerce Clause litigation.
• Sen. Marsha Blackburn's TRUMP AMERICA AI Act, which expands on the White House's four-page framework — markups expected in the Senate Commerce Committee in September 2026.
• Anthropic's next model release and whether it triggers a new Commerce Dept. export-control action — the Mythos precedent makes every frontier release a regulatory event.

The bottom line: Anthropic's state ratchet is a bet that you cannot build two frontier models — one for California and one for the rest of the country. If the company is right, the strictest state's safety law becomes the national standard by default, and Anthropic's competitors pay the cost of catching up to a floor Anthropic has already cleared.