Massachusetts Could Make AI Labs Face Public Risk Tests Every 120 Days

Massachusetts lawmakers are weighing a requirement that major AI labs submit frontier models to independent catastrophic-risk evaluations at least every 120 days, with the findings made public. Anthropic supports the approach; OpenAI says it could delay cybersecurity-model releases and favors a more uniform Illinois-style standard.

The provisions sit inside an economic development bill that cleared the Massachusetts Senate last month. They still need agreement from the House, Senate and Governor Maura Healey. If enacted, the measure would apply only to AI labs with more than $500 million in annual gross revenue, concentrating its immediate obligations on the industry’s largest developers.

A review aimed at catastrophic outcomes

The proposed evaluations go beyond checking whether a company followed its own stated safety process. Outside organizations would assess a model’s potential for catastrophes that kill or seriously injure at least 50 people, or cause $1 billion or more in property damage. The attorney general would set standards for qualifying evaluators, while developers would pay for the reviews.

What the bill would require

• Outside organizations would conduct catastrophic-risk evaluations at least every 120 days, and their findings would be public.
• Labs would disclose safety practices and report safety incidents.
• The state could fine a developer up to $1 million for an initial violation and up to $3 million for later violations involving incident reporting or adherence to its own safety plan.

The dispute is over independence and speed

Anthropic calls the Massachusetts proposal the country’s clearest and strongest AI legislation. Its policy team argues intensive third-party evaluation is necessary because companies should not assess themselves. OpenAI instead backs a uniform standard modeled on Illinois, arguing that inconsistent state rules create confusion rather than added safety.

OpenAI’s more specific objection is operational: it warned that recurring reviews could slow the release of cybersecurity models that might help address the risks lawmakers are trying to manage. The two companies therefore agree on the relevance of external oversight, but differ over its frequency, independence from company rules and possible effect on deployment timelines.

Public visibility without a release veto

The bill’s oversight mechanism is disclosure and compliance enforcement, not direct state control over development. Evaluators could obtain materials reasonably necessary to assess catastrophic risks, but Massachusetts would not gain authority to halt AI development based on their findings. That leaves a central practical question for the final bill: how much leverage public reports and financial penalties will provide when the state cannot block a release.

Negotiators are still working through the AI provisions as part of the larger economic development package. The outcome will determine whether Massachusetts creates a distinct recurring-review regime for large labs or moves closer to the annual audit model OpenAI prefers.