Two governors moved on the same day to impose new controls on artificial intelligence, a coordinated-looking push from states that have decided Washington will not act for them. On September 18, Massachusetts Governor Maura Healey pressed state lawmakers to write independent safety evaluation of the most powerful models into a pending economic bill, while California Governor Gavin Newsom signed an executive order directing state agencies to study a “kill switch” for frontier AI systems.
Healey’s intervention came in the form of a letter to the conference committee negotiating the Mass Wins economic legislation, and it carried a specific demand. She asked the committee to require independent safety assessments of the strongest models before they are deployed, and to force companies to report serious safety incidents more quickly than the current drafts contemplate. The reporting threshold, she argued, should fall from the Senate version’s bar of fifty deaths or a billion dollars in damage to a single death or ten million dollars in property loss.
The threshold is the telling detail. The Senate’s numbers were written to capture catastrophes, the kind of event no governor wants to be the first to confront. Healey’s revision treats even one death as a disaster worth reporting, a standard drawn from the reasoning that a technology this consequential should be governed by its worst plausible outcome rather than its average one. It is a harder standard than the industry has faced, and it is the part of her letter most likely to meet resistance in the committee room.
Newsom’s order, signed the same day, approaches the problem from the other direction. It instructs California agencies to study whether the state can require a kill switch on frontier models, a mechanism to shut a system down if it behaves dangerously, and to draft new safety rules around that capability. The order sets November 16 as the deadline for the study, a tight timeline that signals the governor wants the work done while the technology is still being built rather than after a failure.
California’s path to this point has been uneven. Newsom vetoed a stricter state AI bill, SB 1047, that would have imposed sweeping safety obligations on developers, arguing at the time that it targeted the wrong things and would burden a growing industry. The executive order is widely read as an attempt to recover the ground that veto conceded, offering a narrower, study-first approach that imposes nothing immediately but keeps the question of state regulation alive.
The two states’ moves share a premise: that the federal government, in the current political climate, will not act, and that the states must. Both are acting within the authority they have rather than waiting for a federal framework that shows no sign of arriving, and both are framing their interventions as safety measures rather than as restrictions on innovation, a rhetorical choice meant to blunt the industry’s standard objection.
The industry’s response to state-level regulation has been to warn of a patchwork. Companies have argued that a rule in Massachusetts, another in California, and still more elsewhere would create a compliance burden that differs by border, and that the technology’s risks, being global, demand a single framework rather than fifty. The states’ counterargument, implicit in Healey’s letter and Newsom’s order, is that a patchwork is preferable to nothing, and that the states that act first will define the terms the rest eventually adopt.
Massachusetts and California are natural first movers for different reasons. Massachusetts hosts a dense cluster of AI research in its universities and a growing set of companies, and its governor has made technology governance a theme of her administration. California is home to most of the frontier labs themselves, and a rule that governs the models built in San Francisco and Palo Alto effectively governs the industry, wherever the customers are.
The two approaches also expose a split in how to regulate that the industry will have to navigate. Healey is asking for evaluation and reporting, obligations that attach to a model’s makers before and after deployment. Newsom is asking whether the state can build an off switch, a capability that belongs to the operators of the systems and would require technical means that do not yet clearly exist. One asks the industry to prove its models are safe; the other asks the state to be able to stop them if they are not.
Neither move imposes a binding rule today. Healey’s letter is a request to lawmakers, and Newsom’s order is a study with a deadline. But the timing and the substance together mark a shift in where the center of AI governance is moving. The federal government has made clear it will not be the first to act, and the states, led by two of the country’s most influential governors, have decided they will be.
