The Massachusetts AI safety bill could become one of the strictest state rules on frontier AI models in the country. The language would require independent evaluations of the most powerful models, and it sits inside a broader economic development package that still has no final deal. WBUR reported on October 16 that the proposal could be among the nation's toughest, and it described the bill as still under negotiation.

The AI section belongs to the Mass Wins economic development bill. The Massachusetts Senate passed its version of the package on July 23, 2026, and the Senate press release framed it as an investment in housing, research and responsible AI. The package carries $575.4 million in economic development bond authorizations, so the AI rules are moving alongside a large spending bill rather than on their own.

This is a state story with national spillover, not a new federal action. As of October 16 the package was still being negotiated, and a final deal could land this fall. Nothing in the reviewed coverage shows a conference report or a vote, so treat the current news cycle as renewed attention to a bill still in progress.

What the Massachusetts AI safety bill would require

The Senate's AI section would require covered developers to keep safety frameworks, publish risk reports and hire independent third parties to audit compliance and evaluate their models, according to the New Bedford Guide's report on the Senate draft. The obligations go beyond disclosure. They reach into the testing process itself.

  • In-depth independent safety testing at least every four months, with findings made public. Sources phrase the cadence differently: one says "at least once every 120 days" and another says "at least every four months."
  • Enforcement by the state attorney general, who could sue developers that violate the safety requirements, according to the Boston Globe's summary of the Senate proposal.
  • A catastrophe threshold that evaluators would assess: events that kill or seriously injure 50 people, or cause $1 billion or more in property damage.

The testing schedule matters because it sets a clock. Under the Senate language, outside evaluators would look at a frontier model's risks on a recurring basis rather than during a one-time review, and that recurring structure sits at the center of the independent-review fight.

Which developers would be covered

The Senate's AI section applies to what it calls "large frontier developers," defined as companies with more than $500 million in annual gross revenue [5]. That line points the rules at the largest model builders. It also means the definition turns on revenue, so a company's balance sheet, not only its model, decides whether the law applies to it.

The reviewed coverage centers on frontier developers and does not show direct compliance duties for smaller AI deployers. Businesses that simply use AI tools should still check the final text, since the threshold and definitions could change in negotiation.

Where Gov. Healey and the House fit in

Gov. Maura Healey has pushed to keep the independent-review language. In a Mass.gov release, she urged lawmakers to include independent evaluations of the most powerful models and faster reporting of serious safety incidents in the final bill.

That support has a catch. Healey left the Senate's AI safeguards out of her initial economic development proposal, and both she and House lawmakers still have to agree to the provision [6]. A six-member group of House and Senate negotiators is working out the details, including the AI rules [6].

Timing is the next question. Healey is running for reelection in November, and Insurance Journal reported that a state official expects the bill to pass before the election because it also funds projects in lawmakers' districts [6]. That is a prediction in one account, not a confirmed outcome. As of early October, there had been no announcement of final passage [5].

How Anthropic and OpenAI are responding

Anthropic has endorsed the proposal. Cesar Fernandez, head of U.S. State and Local Government Relations at Anthropic, said the company is endorsing the bill "because it would give Massachusetts the strongest AI safeguards in the country" [2]. In another remark, Fernandez said "we ultimately don't think the industry should grade its own homework" [7] and predicted that other states would follow if Massachusetts adopts these safeguards [7].

OpenAI takes a more conditional position. An OpenAI spokesman said, "We support independent evaluations, including embedded evaluators, and want to help make them effective" [2]. The company wants changes that align the bill with existing state laws [2][7][8]. The Boston Globe reported that OpenAI wants tweaks to match California, New York and Illinois, while Anthropic wants tighter safety measures [8].

Anthropic's push is not limited to Massachusetts. The company has also pitched a federal AI safety bill with narrow blocking power, which suggests it wants a consistent set of rules across levels of government.

How Massachusetts compares with other states and Washington

The Senate proposal goes further than some existing state laws. California, New York and Illinois already have frontier-model safety laws, and Illinois requires a yearly third-party audit [5][6][7]. The Massachusetts testing cadence would be more frequent than that annual audit [5][6], which is the clearest measure of how far the Senate is reaching.

Washington is heading the other way. According to Bloomberg reporting cited by The Next Web, Trump has rejected calls for more AI regulation, arguing it would hold the United States back against China [9]. A federal task force says labs carry responsibility for their models' risks, but no new rules are being written [9].

Congress is also active. Cantwell and Rounds have urged the White House to build a coordinated AI response, a track that runs separately from the state-level push in Boston.

What is still unsettled

  • The final text. Sources describe the negotiations as ongoing and show no conference report or vote [1][5][6].
  • The testing cadence. Sources say "120 days" and "four months," so the final number is not confirmed [5][7].
  • Penalties. Beyond the attorney general's ability to sue, enforcement details are not established in the reviewed sources [8].
  • House acceptance. Healey did not include the Senate language in her original proposal, so House agreement is unresolved [6].
  • The "strongest" label. That description comes from the Senate and Anthropic, not from an independent ranking [5].

What to do now if you build or buy frontier AI in Massachusetts

Ordinary readers have no action to take yet. Professionals who build or buy frontier AI services in Massachusetts can prepare now:

  • Check whether your company's annual gross revenue exceeds $500 million, the threshold for "large frontier developers."
  • Track the final conference text for three things: the revenue threshold, the testing cadence and the attorney general's enforcement language.
  • If you rely on a model provider, ask whether it already runs independent evaluations and publishes risk reports, since those are the duties the Senate draft would place on covered developers.
  • Watch whether House leaders accept the Senate's AI language, because that vote decides whether any of it becomes law.

The next signal will be the negotiators' text. If it keeps the frequent testing schedule and the attorney general's power to sue, Massachusetts would hand other states a detailed template, and Anthropic has already said other states will follow. If House leaders strip the language, the state's push to set the toughest rule will stall, and the federal approach will keep setting the baseline for AI labs.