Massachusetts would force AI safety audits every four months. Anthropic backs it; OpenAI and Google don't.

Massachusetts' state Senate passed an amendment in late July that would make it the strictest state regulator of frontier AI safety on the books -- and the reaction split the three labs the rule would actually bind. Anthropic endorsed the measure back in June, while lawmakers were still drafting it. OpenAI and Google are now lobbying against the version the Senate actually passed. The fight is as much about cadence as substance: the amendment would require the largest AI developers to submit to independent, third-party catastrophic-risk testing roughly every four months, with Massachusetts' Attorney General empowered to sue over violations -- a tighter schedule than even Anthropic described when it first backed the bill.

The provision -- Amendment 471, reported from the Senate Ways and Means Committee on July 16 as part of S.3178, a $325.1 million economic-development bond bill -- applies to what the bill itself calls large frontier developers: companies whose AI systems generate more than $500 million in annual revenue, or that spend more than $1 billion on AI research and development. Those companies would have to write and publish a framework describing how they identify and mitigate catastrophic risks, submit to the roughly-four-month independent testing cycle, protect employees who report safety concerns, and disclose critical safety incidents. Enforcement runs through civil suits the Attorney General's office can bring, not a regulator issuing fines on a schedule.

Anthropic came out for the measure on June 22, the day the Legislature's economic-development committee first released a draft, with Cesar Fernandez -- the company's head of US state and local government relations -- calling it the bill that would give Massachusetts "the strongest AI safeguards in the country." His stated case wasn't about Anthropic being comfortable with scrutiny; it was about who does the scrutinizing, part of what one AI-policy newsletter has already nicknamed Anthropic's state-by-state ratchet on safety rules.

"We ultimately don't think the industry should grade its own homework." -- Cesar Fernandez, Anthropic's head of US state and local government relations

OpenAI disagrees, and hired its first Massachusetts lobbyist -- the firm Benchmark Strategies -- in the weeks before the Senate vote; its opposition was detailed publicly by August 20. Donnie Fowler, OpenAI's head of US state policy, argues that a four-month independent-evaluation cycle amounts to dismantling and rebuilding the engine every few months instead of an annual check, and warns it would specifically slow the release of the defensive cybersecurity models the industry has started shipping this year. His preferred template is Illinois's law: one independent audit a year confirming a company follows its *own* safety guidelines, rather than an outside evaluator assessing the model's actual risk on Massachusetts' clock. Google has taken the same position without a public statement as detailed as OpenAI's.

"Inconsistency doesn't mean safer. It just means confusion." -- Donnie Fowler, OpenAI's head of US state policy

Both companies argue from the same premise -- that a patchwork of state rules is a real cost -- and reach opposite conclusions about which patch to standardize on. California's SB 53, signed in September 2025, and New York's RAISE Act, signed that December 19, were the first state laws requiring frontier labs to publish safety frameworks at all; Illinois followed with annual third-party audits. Massachusetts would be the first to put an outside evaluator, rather than the company's own stated guidelines, in the room, and the first to set the clock in months rather than years. That is the actual disagreement: not whether frontier labs should be scrutinized at all, but who gets to write the test and how often it's given.

That cadence has itself moved while lawmakers argued -- from a roughly six-month cycle in the June draft Anthropic endorsed, to the four-month cycle the Senate actually passed in July -- which is worth stating rather than treating every account of "how often" as interchangeable. (The $500-million-revenue-or-$1-billion-R&D threshold reads broadly, but in practice it currently scopes to a handful of companies -- OpenAI, Anthropic, Google DeepMind, Meta, and arguably xAI -- not the wider AI industry the phrase "frontier developer" might suggest.)

OpenAI's specific worry -- that frequent reviews would slow the release of defensive cybersecurity models -- lands against a pattern this newsroom has already reported: OpenAI, Google and Anthropic each shipped a gated, safety-restricted cyber-capable model within three days of each other at the start of September, every one of them already withheld from open release pending vetted-user programs the companies designed themselves. A Massachusetts-style outside evaluator sitting on top of an access program a lab already built is a genuinely different regulatory question than the one Illinois's annual audit asks -- and it's the one to watch resolve in practice the next time one of those three models updates.

Nothing here is law yet. The Senate passed its version 16-0 out of committee and after two days of floor debate on July 23-24; it still needs to be reconciled with the House's own economic-development bill, passed earlier without the same AI language, before anything reaches Governor Maura Healey's desk. PYMNTS reported in early September that legislative leaders hope to finish that process before November's elections -- a deadline driven by the broader bond bill's other priorities, not the AI fight specifically, which means the frontier-safety language could still be a bargaining chip in a conference committee that has nothing to do with AI at all.