Massachusetts AI Safety Bill: Anthropic vs OpenAI

The fight over how to regulate frontier AI has moved from Washington to Beacon Hill. A Massachusetts proposal to impose some of the strictest state-level AI safety requirements in the country has split the industry’s biggest names: Anthropic is championing the bill’s toughest provisions, while OpenAI and Google are pushing to soften them. The disagreement is narrow on paper — how often frontier developers must submit to independent risk assessments — but it exposes a fault line that will shape AI governance for years.

What the bill requires

The Massachusetts Senate advanced the measure as part of a broader economic development package passed in late July 2026. Its AI provisions target only the largest developers: companies generating more than $500 million in AI-derived revenue or spending over $1 billion on research and development. That threshold is deliberately high, aimed at frontier labs rather than the long tail of startups building on top of their models.

Covered developers would have to publish safety frameworks describing how they identify and mitigate catastrophic risks — the low-probability, high-severity failure modes that safety teams worry about most, from models assisting in the creation of biological or cyber weapons to systems that behave unpredictably at scale. Enforcement would sit with the state attorney general, and the bill kicks off a formal study into independent third-party evaluations of AI systems.

The structure echoes the disclosure-and-transparency approach that has become the template for AI regulation in the absence of federal rules. It does not ban capabilities or gate model releases; it forces developers to say, in public, what they are doing to keep their systems safe — and creates a legal hook if they fail to follow their own stated commitments.

Where the companies split

The dividing line is the cadence of outside scrutiny.

Anthropic is pressing for independent risk assessments every four to six months, a schedule that would make the Massachusetts rules meaningfully more demanding than any comparable state law. The company argues that frontier capabilities are advancing fast enough that an annual review is already stale by the time it is published — a position consistent with its broader posture of favoring binding, enforceable safety commitments over voluntary pledges.

OpenAI supports a narrower version of the legislation and has warned that aggressive state-by-state rulemaking risks creating a patchwork of inconsistent requirements that could delay the release of beneficial models. It favors annual third-party audits modeled on the approach taken in Illinois, rather than the biannual independent evaluations Anthropic wants. In earlier filings, the company cautioned that overlapping and conflicting state mandates would sow confusion for developers operating nationally.

Google has sided with OpenAI in opposing the stricter provisions, putting two of the three largest frontier labs on one side and Anthropic conspicuously alone on the other.

The optics are hard to miss. Anthropic has spent 2026 positioning itself as the safety-forward lab, and reports note the company made campaign contributions exceeding $120,000 to Massachusetts Democrats in August — spending that its critics will read as lobbying and its supporters will read as putting money behind a stated principle.

Why the states are moving

State legislatures are filling a vacuum. Federal AI legislation has stalled repeatedly, and the White House’s own AI safety framework meetings have so far produced convening and guidance rather than binding rules. Into that gap have stepped California, Illinois, New York, and now Massachusetts, each drafting its own definition of what a “frontier developer” owes the public.

For the labs, the nightmare scenario is not any single bill — it is fifty of them, each with different thresholds, reporting formats, and audit schedules. That is the fear OpenAI keeps returning to, and it is a legitimate one: compliance overhead scales with the number of incompatible regimes, and a genuinely conflicting requirement in one state can force a developer to choose which law to break.

The contrast with Europe is instructive. The EU chose a single, continent-wide framework in the AI Act, whose rules for general-purpose models are now entering their enforcement phase. Whatever its flaws, that approach gives developers one rulebook to build against. The United States, absent federal action, is heading toward the opposite — a fragmented map where the rules depend on which state’s users a model serves.

The safety-versus-speed argument

Underneath the procedural fight is a real philosophical disagreement about how to govern a fast-moving technology.

Anthropic’s case is that the pace of capability gains justifies frequent, independent checks — that a model rated safe in the spring may cross a dangerous threshold by the fall, and only a tight review cadence catches that. That instinct mirrors how the whole industry now treats frontier releases — OpenAI, for instance, restricted access to GPT-6 Astra on cyber-risk grounds rather than shipping it wide on day one.

OpenAI’s case is that regulation should be stringent but stable — demanding enough to matter, but predictable enough that developers can plan around it, and uniform enough that a company does not need a different compliance playbook in every jurisdiction. Annual audits, in this view, are not a loophole but a workable equilibrium that captures most of the safety benefit without freezing releases in a permanent review cycle.

Both positions are defensible. What makes Massachusetts consequential is that, for the first time, the disagreement is playing out in a concrete bill with real thresholds, a real enforcer, and real money behind the lobbying on both sides.

What it means

The near-term stakes are small; the precedent is large. Massachusetts is one state, and the bill still has to clear its remaining legislative steps before anything binds. But the version that passes will become a reference point that other states copy or reject, and the four-to-six-month cadence Anthropic is pushing would, if adopted, reset the baseline for what “strict” means in American AI regulation.

Anthropic is making a bet that safety is a durable differentiator. By breaking with OpenAI and Google and spending to back the tougher rules, the company is reinforcing a brand — the lab that welcomes scrutiny — that matters to enterprise buyers and to regulators alike. It is also a bet that arrives as Anthropic prepares for one of the most closely watched IPOs in tech history, where a clean regulatory story is an asset.

Watch the enforcement details, not the headlines. Publication-of-frameworks requirements are only as strong as the attorney general’s willingness to act on them, and audit-cadence rules only bite if the independent evaluators exist and have access. The study the bill commissions into third-party evaluation is, quietly, the most important part — because a serious independent audit ecosystem is the thing every version of this debate ultimately depends on.

And watch whether the patchwork forms. If Massachusetts, California, and Illinois end up with three genuinely incompatible regimes, OpenAI’s warning becomes self-fulfilling and the pressure for a single federal standard grows. If they converge on something close to a common template, the states will have written the rules Washington could not.