Massachusetts Nears Toughest State AI Safety Law, Setting Up a Fight With Trump's DOJ

Massachusetts lawmakers are in the final stretch of negotiating what could become the country's most aggressive state AI safety statute, and the fight is no longer abstract: it is now a live lobbying battle between Anthropic on one side and Google and OpenAI on the other, playing out inside a conference committee that is supposed to reconcile the state's economic development bill by the end of this month.

The disputed language, pushed by Gov. Maura Healey and state Sen. Barry Finegold, would require independent evaluators to test the most powerful AI models for "catastrophic risk" — defined in the bill as an event causing death or injury to 50 or more people or $1 billion or more in damage — at least three times a year. The rules would apply only to the largest developers, those with more than $500 million in annual revenue and the computing scale to train frontier systems, and the companies themselves would have to pay the evaluators. Violations would carry penalties of up to $1 million for a first offense and up to $3 million for repeat violations.

Three existing state laws, and a gap none of them close

California, Illinois and New York already require AI developers to publish safety plans, report serious incidents, or protect whistleblowers. Illinois goes furthest among them, mandating annual third-party audits. But none of those laws let outside evaluators actually examine how a model behaves before or after release — the access gap the Massachusetts proposal is designed to close, and the reason state officials are describing it as a meaningfully different category of state law rather than a variant of what already exists elsewhere.

That distinction is exactly what has split the industry. Anthropic has publicly endorsed the Massachusetts language, with its government relations lead Cesar Fernandez telling reporters the company is "endorsing this bill because it would give Massachusetts the strongest AI safeguards in the country." Anthropic has separately announced plans to bring safety evaluators in-house through a partnership with Accenture, a move that dovetails with — and may commercially benefit from — a regulatory regime built around recurring third-party testing.

OpenAI and Google are pushing the other direction. In letters sent to the conference committee and to Finegold in August, OpenAI argued Massachusetts should instead mirror Illinois's lighter, annual-audit model, and has since said it wants independent assessments handled at the federal level rather than state by state. A company spokesperson said OpenAI "supports independent evaluations, including embedded evaluators, and wants to help make them effective," without endorsing the specific testing cadence or outside-access provisions in the Massachusetts bill. Google has argued separately that no agreed testing standards yet exist, leaving compliance and results open to dispute. Meta, Google, Anthropic and OpenAI are all actively lobbying the legislature, according to Finegold; Meta, Google and Amazon did not respond to reporters' requests for comment on the bill itself.

A law built to collide with Washington

The timing raises the stakes beyond Massachusetts. A December 2025 executive order from President Trump directed the Department of Justice to stand up an AI Litigation Task Force aimed at challenging state AI laws the administration deems "onerous," part of a broader push for industry self-regulation over a state-by-state patchwork. Trump has been blunt about his preferred approach, posting that "the only control or 'guardrails' that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT." If Massachusetts enacts mandatory third-party model testing with real penalties attached, it would be a direct test of how far that task force is willing to go against a state law that has bipartisan executive backing — Healey has said companies building these systems "have a responsibility to build them safely" — rather than the ballot-driven or legislature-only statutes the administration has criticized elsewhere.

Critics inside Massachusetts are not only worried about Washington. Tech investor Paul English has asked what the state gains by regulating frontier models if Chinese developers face no equivalent obligation, and analysts note that open-weight models released publicly — including by Chinese labs — can be modified well outside the reach of any state evaluator regime once they leave the original developer's hands. Separately, Paul Baier, CEO of GAI Insights, has claimed Anthropic and OpenAI privately favor scaling penalties to one to five percent of a company's market value rather than the flat dollar caps in the current draft, a structure that would scale punishment to company size rather than capping it outright.

What happens next

The evaluator provision's fate now rests with the conference committee reconciling the Mass Wins economic development package, with Finegold saying as of this past week that the issue was "still being discussed" and final language expected within weeks. The outcome will matter well beyond Massachusetts: it is the first state proposal to mandate recurring, paid, outside testing of frontier models for catastrophic risk rather than self-reported safety plans or periodic audits, and a signed law would hand the Trump administration's litigation task force its clearest opportunity yet to test federal preemption against a state AI statute with real teeth.

Cesar Fernandez, Government Relations Lead, Anthropic: "We're endorsing this bill because it would give Massachusetts the strongest AI safeguards in the country."