Negotiations in the Massachusetts legislature could bring strictest AI safety law in the US
A proposed law in the state Senate's economic development package would mandate independent catastrophic risk evaluations for powerful AI models at least three times a year.
A new economic development bill being negotiated in the Massachusetts legislature could see the state adopt the strictest AI safety regulations in the United States. The law, part of the state Senate’s version of the bill, would require independent evaluators to test powerful AI systems for catastrophic risks at least three times a year — a mandate that goes beyond laws already enacted in California, Illinois, and New York.
The provision would mark one of the country’s most forceful attempts to address what researchers call frontier AI risk: the potential for highly capable general-purpose models to cause large-scale harm through misuse, accident, or loss of control. Where other state laws have focused on sectoral use cases or discrimination audits, the Massachusetts approach would mandate direct third-party evaluation of the core model itself for those catastrophic capabilities, regardless of the specific application. Independent evaluators would be required to examine the most powerful models at least three times annually, testing for dangers that could include large-scale disinformation, cyberattacks, or autonomous threats that even developers might not fully anticipate during training.
The measure is not yet law. It is part of ongoing bicameral negotiations between the state House and Senate over the broader economic development package, meaning the exact language could still change or be removed before a final vote. But the fact that it has advanced to this stage in the Senate’s bill signals a serious legislative push toward treating AI not just as an economic asset but as a technology whose most advanced forms may need pre-deployment safety checks modeled more on pharmaceuticals or aviation than on consumer software.
If it survives negotiations and passes, Massachusetts could adopt some of the nation’s toughest artificial intelligence regulations by the end of the current legislative session. That would position the state — home to a dense cluster of AI research labs, academic institutions, and tech companies — as a testing ground for a regulatory framework that other governments, including the federal one, have so far approached more cautiously. The requirement for regular, independent catastrophic-risk evaluation would set a concrete compliance burden that even the EU’s AI Act, for all its breadth, does not impose at the same frequency or with the same focus on frontier-model internals.
What remains unanswered is how the state would define which models are “powerful” enough to trigger the evaluation mandate, who would qualify as an independent evaluator, and what standards they would use to assess a risk that by definition has not yet materialized. The bill’s proponents appear to be betting that the mere requirement to conduct and disclose these assessments will force developers to build safer systems and give regulators a clearer window into capabilities that are currently opaque even to their creators. Whether that bet pays off — and whether it sets a template other states follow or becomes an outlier that pushes AI development elsewhere — will depend on the final text, and on whether the companies that would be most affected decide to comply, challenge, or simply avoid the jurisdiction altogether.