Massachusetts Wants Independent Audits of Frontier Models Every Four Months, and the Labs Disagree About It
Anthropic backs a state bill requiring outside evaluators to check large models for catastrophic risk three times a year. OpenAI and Google want something closer to Illinois's annual audit, warning about a patchwork of state rules.
A bill moving through the Massachusetts Senate would require large AI developers to hire independent evaluators to assess their models for catastrophic risk every four months. The unusual part is not the rule. It is that the industry has split publicly over it: Anthropic is for it, OpenAI and Google are against.
The AI language sits inside a broader economic development bill. Anthropic’s head of US state and local government relations, Cesar Fernandez, wrote when the Senate passed its version that it “raises the bar on AI safety while ensuring that innovation is able to continue.” Safety organisations including Encode and the Secure AI Project have sent letters of support to the conference committee.
OpenAI wrote to lawmakers last month arguing for something closer to an Illinois bill, which asks for annual third-party safety audits rather than three per year. Its objection is less about the audit itself and more about arithmetic: if every state writes its own version, developers face what OpenAI called “fragmented oversight” and an “unhappy result.”
What is behind this
An independent evaluation here means an outside organisation, not the lab’s own safety team, testing a model for the worst-case capabilities: helping build weapons, running large-scale cyberattacks, that category. Doing it every four months rather than annually matters because frontier models now ship faster than that. An annual audit can miss two model generations.
The states are moving because Washington is not. Congress rejected a proposed ten-year moratorium on state AI laws in 2025, and a later attempt to attach preemption language to defence legislation also failed. Rhode Island enacted chatbot safeguards in June, California is working on audits and privacy, and Utah and South Carolina are back on children’s online safety. Massachusetts would be the strictest of the lot.
It also helps to notice the commercial angle. Anthropic has built its market position on being the safety-forward lab, so rules that make third-party evaluation mandatory raise a cost its competitors have not been paying and it largely has. That does not make its position insincere. It does mean the split is not purely a disagreement about ethics.
What this means for you: Nothing you will notice this month. But state rules are where AI regulation in the US is actually happening, and the cadence Massachusetts picks will probably get copied. For anyone building on these models, a four-month audit cycle means the safety documentation you rely on gets refreshed more often, which is genuinely useful. For everyone else, this is worth watching as the first case of the big labs openly lobbying against each other over safety rules.
Sources
The US Justice Department Tells a Court That Training AI on Copyrighted Books Is Fair Use
In the consolidated New York Times case against OpenAI and Microsoft, the DOJ filed on the side of the AI companies and directly contradicted the US Copyright Office report whose author was fired last year.
