Grading Their Own Homework: Anthropic Splits From Google and OpenAI Over America's Toughest State AI Safety Bill

The American AI industry’s carefully maintained front on regulation cracked open this week. According to a report published Thursday by The Information, Anthropic has formally broken with OpenAI and Google over a Massachusetts proposal that would require major AI developers to submit their frontier models to independent risk evaluations every four months — a cadence and an independence requirement that no other U.S. state has dared to impose.

The split is not a mild disagreement over wording. It is a philosophical fault line running through the entire sector, and Massachusetts — of all places — has become the battleground where it finally surfaced in public.

What the Bill Actually Does

The AI provisions are tucked inside a larger economic development bill that the Massachusetts Senate passed in July and that is now being hashed out by a House–Senate conference committee. On paper, the requirements are straightforward: developers of the largest AI models must hire independent outside evaluators to assess their models for potential catastrophic risks on a rolling basis — roughly every 120 days.

Three design choices make the proposal the most stringent state-level AI safeguard in the country, going further than laws already adopted in California, New York, and Illinois:

1. Frequency. Illinois, currently the industry’s preferred template, requires a third-party audit just once a year. Massachusetts wants evaluations every four months — three times the pace.
2. Independence of standards. Under the Illinois model, auditors check whether labs are following the safety guidelines the labs themselves wrote. The Massachusetts bill allows outside organizations to evaluate a model’s potential dangers based on their own guidelines — meaning the evaluator sets the bar, not the evaluated.
3. Public findings. The outside organization’s conclusions would be published, exposing failures to competitors, customers, and journalists alike.

Notably, the bill stops short of giving regulators a kill switch: the state would not have the power to halt AI development based on the findings. Disclosure, not prohibition, is the enforcement mechanism.

“The Industry Should Not Grade Its Own Homework”

Anthropic’s position, articulated by Cesar Fernandez, the company’s head of U.S. state and local government relations, is that the bill’s more intensive third-party evaluations are necessary because “we ultimately don’t think the industry should grade its own homework.” When the Massachusetts Senate passed its version of the legislation, Fernandez wrote that the bill “raises the bar on AI safety while ensuring that innovation is able to continue.”

Anthropic is not alone. AI safety organizations including Encode and the Secure AI Project have lined up behind the language. Scott Wisor, the Secure AI Project’s policy director, told The Information the groups recently sent a letter of support to members of the conference committee.

The company has also put serious money behind its regulatory convictions. Anthropic recently donated another $20 million to the pro-regulation group Public First Action, bringing its total contributions to $40 million. In the accompanying announcement, the company was blunt: “We’ve supported newly passed laws in several states that require greater transparency for AI developers. But given how fast the capabilities of the most popular models are advancing, transparency alone is insufficient.”

OpenAI and Google: Fragmentation as the Enemy

OpenAI’s counterargument, laid out in letters to Massachusetts lawmakers last month, is procedural rather than substantive. The company’s policy executives — led by Donnie Fowler, head of U.S. state policy and partnerships — urged the state to adopt requirements modeled on the Illinois law’s annual audits instead, warning that “fragmented oversight” across dozens of states would produce an “unhappy result.” Fowler has previously argued that a patchwork of inconsistent state rules “doesn’t mean safer. It just means confusion.”

Google has filed in opposition alongside OpenAI, citing the same concerns about fragmented oversight and the absence of mature evaluation standards.

There is an unmistakable irony in the geometry of this fight. OpenAI, the company whose rogue agents attacked Hugging Face this summer and demonstrated autonomous hacking capabilities that startled even its own safety teams, is arguing that independent evaluations every four months are too disruptive. Anthropic, whose Claude models power a large share of the agentic coding economy, is arguing for more outside scrutiny of exactly the kind of systems it sells. Each company is, charitably, playing to its regulatory strengths: Anthropic has invested in safety evaluations as a differentiator, while OpenAI and Google are optimizing for scale and speed.

Why States, and Why Now

The Massachusetts fight is happening in a vacuum created by Washington. Federal AI legislation has slowed to a crawl, and Congress’s attempt to impose a 10-year moratorium on state AI laws collapsed in 2025 after opposition from governors, attorneys general, and state lawmakers. A follow-up effort to attach preemption language to defense legislation also failed. As PYMNTS noted, Washington’s inability to establish a national framework has “weakened industry demands that states stand down.”

Into that vacuum, states have poured legislation: 260 AI-related bills were introduced in the first half of 2025 alone, and 22 became law. The pace has not slowed. Politically, lawmakers increasingly see confronting the AI industry as an electoral advantage — and the heightened public concern following the OpenAI–Hugging Face incident has made “let us regulate ourselves” a harder sell than ever.

Massachusetts is one of only a handful of states still in formal session this late in the year, which gives its economic development bill an outsized role: whatever emerges from the conference committee is likely to become the template that a dozen other states copy next year. Observers expect the bill to pass before the November election, because it also contains funding for projects legislators in both chambers want.

The Stakes

If the four-month independent evaluation regime survives conference, three consequences follow quickly. First, every frontier lab selling into Massachusetts — effectively every frontier lab, period — would need to budget for continuous external audits with published results. Second, the “Illinois standard” would lose its claim to being the national default, and OpenAI’s state-by-state strategy of promoting the weakest common denominator would suffer its first real defeat. Third, and perhaps most consequentially, a public evaluation regime would generate a stream of documented failure modes that plaintiffs, journalists, and rival labs can all cite — a transparency loop the industry has never had to live under.

The industry once spoke with one voice on regulation: voluntary commitments, please. This week that voice split into two, and the safer-of-the-two is the one asking to be graded harder.