New York's Next AI Bill Would Let Regulators Order Labs to Throttle Their Own Training Runs
State Senator Andrew Gounardes spent most of 2025 fighting to pass New York's RAISE Act, then watched Governor Kathy Hochul spend the first half of 2026 negotiating it down toward something friendlier to industry. On October 28, he introduced a bill that reads like an attempt to take the law back past where it started. S10731, co-sponsored with Assemblymember Jordan Wright, would give the state's new AI oversight office the explicit power to order a frontier AI developer to limit the computing power it uses to train a model or run automated AI research. Gounardes isn't being subtle about the intent: in the bill's rollout, he said it "creates safety and transparency requirements with real teeth," and framed the push bluntly as a need to "rein in the AI arms race."
From disclosure regime to stop-work authority
The RAISE Act that Hochul actually signed in December 2025 was, at bottom, a reporting law. Large frontier developers have to publish safety frameworks, disclose critical incidents to regulators within 72 hours, and register with a new Department of Financial Services unit called DIGIT — the Office of Digital Innovation, Governance, Integrity and Trust — ahead of a registration window that opens this November and a compliance deadline of January 1, 2027. When industry groups pushed back hard last winter, Hochul's office negotiated amendments that pulled the law closer to California's SB 53, which relies on transparency and disclosure rather than giving regulators direct control over how a lab trains its models. Assemblymember Alex Bores, who helped write the original law, has said publicly that DIGIT was always meant to be able to "recommend new legislation as the technology evolves" — and this bill is that recommendation arriving less than a year later, before the law it amends has even taken effect.
S10731 would add a new section letting DIGIT order compute limits on a covered developer — not just ask for a safety plan, but actually cap the resources behind a training run or an AI R&D process the office judges too risky. It also lowers the compute threshold that defines a "frontier model" from 10^26 to 10^25 floating-point operations, which sweeps more systems into coverage, and it would phase in state-licensed third-party "frontier examiners" by 2029 to conduct audits DIGIT itself doesn't have the staff to run. Penalties would climb to as much as $1 million per day for a first violation and $3 million per day for repeat offenses, and the attorney general could seek a court order restricting a model's deployment inside New York altogether.
A kill switch by another name
The mechanism is notable because it's close to the one piece of AI safety law that has consistently failed to survive contact with a governor's desk. California's SB 1047, which would have let regulators demand a "full shutdown" capability for the most powerful models, was vetoed by Gavin Newsom in 2024 largely on the argument that it reached too far into how companies actually build and run their systems — as opposed to what they disclose about them. New York's RAISE Act, both in its original and amended forms, deliberately avoided that kind of operational authority. S10731 reintroduces it through a different door: instead of a blanket shutdown mandate, it gives an administrative agency discretionary power to limit compute case by case, which may prove more durable in court even though the practical effect — a state telling a company how much it's allowed to compute — is similar. Hochul herself floated the idea of "AI kill switches" as a preliminary concept in a September 21 announcement about DIGIT's buildout, calling it something the state "may even explore" without commitment. Two months later, her own RAISE sponsor has put a concrete version of that idea into bill text.
The bill also expands whistleblower protections for employees at covered AI companies and bars contracts that would silence them — a provision Wright pitched directly to workers inside the labs: "When folks inside one of these companies see something wrong and speak up, New York will have their back." Encode's Ben Snyder, whose group lobbies for stronger AI safety rules, backed the bill on the grounds that it would help ensure "people — not profits or the pressure of the race to the top — are driving AI development," language that doubles as an acknowledgment that the current law, post-amendment, hasn't fully delivered on that front.
Timing that invites a fight
The bill lands at an awkward moment for anyone hoping state AI law might cool off. The Trump administration has spent months building a legal campaign — including a Justice Department task force created by executive order — aimed at getting courts to strike down state AI statutes as obstacles to a single national framework. A law that lets a state agency order a company to throttle the compute behind a training run is a considerably more attractive target for that campaign than one that merely requires a safety memo, since it reaches further into the operational and interstate aspects of how AI companies actually run their business. Industry groups that lobbied Hochul into softening the original RAISE Act are likely to treat this bill as the fight they thought they'd already won a version of.
For now, S10731 is sitting in the Senate Rules Committee, with the legislature not due back in regular session until January — which means nothing in it will matter before the RAISE Act's own November registration deadline arrives. But its introduction answers a question that's been hanging over Albany since Hochul's concessions last winter: whether New York's appetite for binding AI oversight survived the compromise. Gounardes and his allies are betting that Bores's framing — the RAISE Act as "the floor, not the ceiling" — still has enough support to build something taller on top of it, even as Washington tries to tear the floor out from under all of it.