Analysis
Massachusetts lawmakers are negotiating what would be the strictest state-level AI safety law in the country, and the fight over it has done something unusual: split the three biggest frontier AI labs publicly, in writing, on the same bill. The Massachusetts Senate's economic development committee released the provision on June 22 as part of a $325 million economic development bond bill, and Anthropic announced its endorsement four days later, calling it the strongest AI safety legislation of any US state so far. OpenAI and Google have both come out against the specific provisions Anthropic is backing, and as of early September the bill sits in bicameral conference committee, meaning the final language is still being negotiated between the House and Senate versions.
The mechanism at the center of the fight is a requirement that frontier labs generating more than $500 million in AI-derived revenue, or spending more than $1 billion on AI research and development, submit to independent third-party reviews of the catastrophic risks their models pose -- on a cadence Anthropic itself is reportedly pushing to tighten to every four to six months. That threshold is drawn narrowly enough to catch only a handful of companies: OpenAI, Google, Anthropic, and likely Meta and xAI as their AI businesses scale, but not the thousands of startups building applications on top of their APIs.
Where the labs actually disagree
- Anthropic: publicly endorsed the Massachusetts bill on June 26 and is reportedly pushing for reviews as frequent as every four months, arguing that Massachusetts should set the strictest bar in the country given the pace of frontier capability gains.
- [OpenAI](/pulse/company/openai): opposes the bill's specific approach. Donnie Fowler, OpenAI's head of U.S. state policy and partnerships, has argued publicly that other states should model their standards on Illinois's lighter-touch law, saying regulatory inconsistency across states "doesn't mean safer, it just means confusion."
- Google: has also opposed the stricter Massachusetts approach, aligning with OpenAI's preference for a more permissive, Illinois-style national baseline rather than a state-by-state patchwork with Massachusetts setting the ceiling.
That three-way split matters because it undercuts the industry's usual negotiating posture with state legislatures, which has generally been a unified push for federal preemption over any state-level patchwork. At the federal level, the current administration has pushed for a looser national approach and moved in December 2025 to block state AI laws flagged as incompatible with a minimally burdensome federal framework -- a policy that would seem to cut directly against what Anthropic is now backing in Massachusetts. Nearly 100 chatbot-specific AI bills have been introduced across 34 states in 2026, and Connecticut, Texas and Illinois have all passed their own versions this year, but Massachusetts's proposed 120-day catastrophic-risk review cadence would be the first mechanism of its kind for large frontier labs specifically, rather than for chatbot-facing consumer products.
The counterweight worth naming here: legislative language and enacted law are not the same thing, and this bill is still in conference committee with an outcome that could land anywhere between the Senate's strict version and something closer to what OpenAI and Google are asking for. A state law, even a strict one, also only binds companies with a Massachusetts nexus for compliance purposes -- it does not by itself create a national standard, and a company could in theory structure its compliance operations to minimize exposure the same way OpenAI's preferred Illinois-modeled approach would allow.
For founders building on top of frontier models, the practical question is less about whether their own startup gets swept into the $500 million revenue threshold -- almost none will be, for years -- and more about which compliance regime their model provider ends up operating under, since a 120-day catastrophic-risk review cycle could plausibly slow release cadence or gate certain capabilities behind additional review for the labs it applies to. Watch whether the conference committee's final language keeps Anthropic's preferred four-month cadence or drifts toward OpenAI and Google's looser preference -- that outcome will tell you more about which lab actually has more regulatory influence in state capitals than any public statement has so far.
