Anthropic, OpenAI Clash Over Strict Massachusetts AI Safeguards

Massachusetts lawmakers are advancing a bill that would require leading AI labs to undergo independent reviews of catastrophic risks every 120 days, a first for a US state, drawing a split between Anthropic PBC, which supports the measure as 'the clearest and strongest AI legislation in the country,' and OpenAI, which warns the reviews will slow cybersecurity model releases and prefers a uniform standard like Illinois' annual audit. The provision, part of an economic development bill passed by the state Senate, goes further than laws in California, New York, and Illinois, and must still be approved by Governor Maura Healey and House lawmakers.

Bloomberg -- Massachusetts lawmakers are seeking to establish the nation's most stringent state-level safeguards for artificial intelligence, a proposal that's divided an industry the governor's office is concurrently seeking to woo. Most Read from Bloomberg An economic development bill passed by the state Senate last month included a requirement that leading AI labs undergo independent reviews of the catastrophic risks posed by their frontier models at least once every 120 days, a first for a US state. The findings would be made public, though the state couldn't use them to halt AI development. The provisions go further than other recent laws in California, New York and Illinois and come as Massachusetts Governor Maura Healey and other leaders seek to build up the region's AI sector and curb the outflow of startup talent. Healey, a Democrat running for reelection in November, didn't include the AI safeguards passed by the Senate in her initial economic development proposal. Both she and House lawmakers still must agree to the provision. Leading AI developer Anthropic PBC has advocated for states to adopt more stringent AI regulation in the absence of a federal standard and calls the Massachusetts proposal "the clearest and strongest AI legislation in the country." OpenAI, on the other hand, is warning that the reviews will slow the release of cybersecurity models that might protect against the very risks lawmakers fear and prefers states adopt a uniform standard in line with the Illinois law. Inconsistency "doesn't mean safer. It just means confusion," said Donnie Fowler, OpenAI's head of US state policy and partnerships. While Illinois now requires a third-party audit, it's only once a year and geared toward ensuring labs are following their own safety guidelines and making disclosures mandated under the law. The Massachusetts proposal includes a similar annual compliance review, but separately empowers outside organizations to make evaluations of a model's dangers, independent of the developer's guidelines, every few months. The evaluators would assess the potential for catastrophes that kill or seriously injure at least 50 people or lead to $1 billion or more in property damage. Fowler likens Illinois' audits to an annual car inspection, the equivalent of making sure the brake lights and windshield wipers are working. Massachusetts' more frequent evaluations would be akin to dismantling the engine and then putting it back together again, he said. Anthropic believes the more intensive third-party evaluations are needed because "we ultimately don't think the industry should grade its own homework," said Cesar Fernandez, the company's head of US state and local government relations. While Anthropic endorses the Massachusetts proposal, states may need to adopt even tighter oversight in the future as the technology advances, he said. Anthropic and OpenAI are jostling to shape the final legislation. Anthropic hired Boston lobbying firm Tremont Strategies Group earlier this year, while OpenAI brought on local firm Benchmark Strategies as its lobbyist in recent weeks. The broader bill is likely to pass before the November election because it also contains funding for projects in lawmakers' districts. A six-member group of House and Senate negotiators are working to iron out the details, including the AI rules. State Senator Barry Finegold, one of the negotiators, said independent evaluations are needed to ensure frontier models won't cause significant harm. If Massachusetts adopts the safeguards, other states will follow, he said. "We were behind on social media. We allowed too much to go by without putting in guardrails, and now we're playing catch-up. We don't want to do that again," Finegold said. AI safety has taken on more urgency in the past month after OpenAI, Anthropic and Meta Platforms Inc. each reported instances in which their models inadvertently hacked into third-party systems in testing environments. Federal oversight is limited: President Donald Trump has taken a largely hands-off approach and bipartisan AI safety legislation that would preempt state laws is stalled in Congress. States are starting to fill the gap. At the same time, they're also competing with each other for a piece of the AI boom that's driving an outsize portion of the country's current economic growth. In February, Massachusetts announced a partnership with OpenAI and became the first state to give its entire executive branch access to a ChatGPT-powered AI assistant. Early this spring, Healey and her economic development secretary, former venture capital executive Eric Paley, spoke with Anthropic President Daniela Amodei about the company potentially growing its Massachusetts presence and partnering with the state in areas like cybersecurity and workforce development, according to emails obtained through a public records request. Fernandez, the Anthropic government relations executive, invited Healey and Paley to its San Francisco headquarters, in part to discuss "responsible governance frameworks that serve as a national model." In a follow-up message, Paley suggested other partnerships to explore, including on biotechnology and AI literacy in schools. He told Fernandez that "we all need to share more clear examples of how AI is enhancing humanity and offer positive inspiration to combat all of the fear mongering." Paley offered to reach out to a state legislator about collaborating on AI oversight. The Healey administration will work with lawmakers on finalizing the legislation this fall, with a goal of setting "appropriate safeguards on AI development while continuing to support our innovation economy, jobs and competitiveness," a spokesperson for the state's economic development arm said. The topic is likely to feature in the upcoming governor's race. Michael Minogue, a former medical device executive who's the leading contender for the Republican nomination, dismissed the proposed third-party evaluations as overregulation. "How would a non-expert advise an expert building something every 120 days?" Minogue said in an interview at Bloomberg's Boston office last week. "If you're an AI company, I'm not coming here now." Under the Massachusetts proposal, evaluators would have the ability to interrogate AI labs and obtain access to "all materials reasonably necessary" to assess catastrophic risks. The attorney general would create standards for outside organizations to serve as evaluators. The bill allocates no government funding for this task, instead allowing developers to pay evaluators. Only AI labs with annual gross revenues over $500 million would be subject to the evaluations. Like the California, New York and Illinois laws, the Massachusetts measure would also establish whistleblower protections for AI-lab employees and require developers to disclose safety practices and report safety incidents. Should an AI developer fail to report an incident or adhere to its own safety plan, the attorney general could pursue a fine of as much as $1 million for an initial violation and as much as $3 million for additional ones. The proposed safeguards strike the right balance by keeping officials and the public in the know about risks without compromising the AI industry's long-term success, said Ben Snyder, a policy adviser at Encode, an advocate for AI safety limits. "The risk assessments are really essential to helping governments keep up and have real conversations with companies," he said. Most Read from Bloomberg Businessweek ©2026 Bloomberg L.P.