Summary: With federal AI legislation stalled, Anthropic has pursued state-by-state bills that progressively raise safety requirements, while OpenAI promotes a California-style template across states. The article contrasts Anthropic’s opposition to Illinois SB 3444’s broad liability shield with both companies supporting SB 3261, which requires annual independent third-party audits of safety plans. It says the winning approach may shape future federal legislative drafts.
With Washington unable to move on AI legislation, the fight has shifted to the statehouses — and that is where America’s two leading frontier labs have publicly parted ways. Wired and Business Insider reporting shows Anthropic and OpenAI giving incompatible answers to what a state AI safety law ought to look like, and nearly every state-level fight since the summer of 2026 has run along that dividing line.
Two Strategies: Ratcheting Up vs One Template Everywhere
Anthropic is playing a ratchet game: it backs ever-tougher bills, one state at a time, so that strict standards push each other higher. State policy lead Cesar Fernandez puts it this way: “We’re looking for legislation that meaningfully raises the bar on safety,” adding that action “can’t wait for action in Washington.” OpenAI’s approach, steered by its top lobbyist Chris Lehane, is “reverse federalism”: copy one California-style framework into state after state until it amounts to a national standard in practice. Spokesperson Liz Bourgeois argues that this “gives the public clearer protections” and keeps developers from facing conflicting rules.
Illinois SB 3444: A Head-On Fight Over the Liability Shield
Illinois is where the split cuts deepest. SB 3444, introduced by Senator Bill Cunningham, would exempt labs from liability when their systems are used to cause large-scale harm — mass casualties or more than $1 billion in property damage — provided the lab had drafted and published a safety framework on its own website. OpenAI is behind the bill, saying it lowers serious-harm risk while still letting the technology reach Illinois businesses.
Anthropic lobbied hard to change the bill or stop it. Fernandez did not mince words: transparency laws should “ensure public safety and accountability,” not hand out “a get-out-of-jail-free card against all liability.” Governor JB Pritzker is against it as well — he “does not believe big tech companies should ever be given a full shield.” Thomas Woodside of the Secure AI Project said the bill “would take the extreme step of nearly eliminating liability for severe harms.” Policy experts see the odds of SB 3444 passing as remote, yet the episode laid bare the political distance between the two companies.
Where the States Stand
- California: the 2025 law — the first state law in the country to regulate advanced models — drew an endorsement from Anthropic and no other major lab that year; Anthropic treats it as a floor, not a ceiling
- New York: as Business Insider reported, the pressure on Governor Hochul late last year to move the state’s bill toward the California framework came from OpenAI’s lobbyists, not Anthropic’s; Anthropic had backed the stronger original text
- Illinois: Pritzker signed the audit-direction bill (SB 3261) in July 2026, requiring safety plans to undergo annual independent third-party audits — the first mandate of its kind, backed by both OpenAI and Anthropic
- Massachusetts: in late June, Anthropic endorsed provisions in an economic development bond bill, calling them the toughest in the nation — independent evaluators assessing catastrophic risks (such as bioweapons assistance), with enforcement by the state attorney general; OpenAI is still reviewing
Money and Motives
The state-level campaign is getting costly: dueling super PACs aligned with each company have spent tens of millions, and since June Anthropic has been donating directly to California legislators. Critics — some of them close to the Trump administration and to venture capital — cast Anthropic’s drive as regulatory capture meant to hobble competitors.
Business Insider also flags a technical motive: internal Anthropic testing found that Claude Mythos could exploit vulnerabilities in every major operating system, and the public sibling model Fable triggered Trump administration export controls. In effect, Anthropic is putting its own strongest model’s test results forward as first-hand evidence for why strict rules are necessary.
What to Watch
Teams following AI governance should track two things: whether other states copy Illinois’s third-party audit model (so far the ratchet strategy’s most concrete result), and whether OpenAI’s uniform template can claim its first win in a large state. Whichever approach prevails will end up as the draft text of future federal legislation.