New York's RAISE Act: a second American frontier-AI safety law

While Washington debates whether to regulate artificial intelligence at all, a second American state has decided it for itself. On 27 March 2026 Governor Kathy Hochul signed the chapter amendments that finalised New York's Responsible AI Safety and Education Act (the RAISE Act, Senate Bill S6953B / Assembly Bill A6453B), completing a law she had first signed on 19 December 2025 on the condition that the legislature pass agreed changes. The Act takes effect on 1 January 2027. With it, New York becomes the second state after California to write a safety statute aimed not at everyday AI use but at the handful of companies training the most powerful models — and because those companies supply the systems used in Europe and worldwide, this is not a development that organisations outside the United States can treat as purely domestic.

Who the Act reaches

The RAISE Act, like California's frontier-AI law, regulates a narrow tier defined by raw training compute. A frontier model is an AI model trained using more than 10²⁶ computational operations, the compute cost of which exceeds 100 million dollars. The heaviest duties fall on the large developer: a person that has trained at least one frontier model and whose gross annual revenue exceeded 500 million dollars in the last calendar year.

Both numbers were chosen deliberately to track California. The 10²⁶ operations trigger is identical to the SB 53 threshold, and it sits an order of magnitude above the 10²⁵ FLOP line the EU draws for general-purpose AI models with systemic risk — New York, like California, aims at a still narrower band than Brussels. The 500-million-dollar revenue test is the same one California uses to identify its largest developers. Governor Hochul framed the law explicitly as building on "California's recently adopted framework."

What the Act requires

The obligations rest on three pillars, deliberately close to California's.

First, a safety and security protocol. Large developers must write, implement and conspicuously publish — with appropriate redactions — a protocol describing how they identify, assess and mitigate the risk that a frontier model causes catastrophic harm, including their cybersecurity practices, governance and the thresholds at which they would decline to deploy or would shut a model down. The protocol must also be transmitted to the New York Attorney General and the Division of Homeland Security and Emergency Services.

Second, safety-incident reporting. A large developer must disclose a critical safety incident within 72 hours of learning of it — or of learning facts that establish a reasonable belief that one occurred — and within 24 hours where the incident poses an imminent risk of death or serious physical injury.

Third, whistleblower protection. The Act bars developers from retaliating against employees who disclose information about catastrophic risk or about violations of the law to the Attorney General or other authorities.

Catastrophic risk, not everyday decisions

Like SB 53, the RAISE Act reaches past ordinary harm to the tail risks of the most capable systems: a frontier model materially enabling the creation of a chemical, biological, radiological or nuclear weapon, or acting without meaningful human oversight in a way that causes mass casualties or large-scale damage. This is a different regulatory object from the state employment and consumer laws — Colorado's reworked automated-decision-making statute and the others — that govern hiring, credit and algorithmic discrimination. New York now operates on both planes: a frontier-safety regime here, and ordinary AI consumer protections elsewhere in its law.

Enforcement and oversight

Enforcement sits with the New York Attorney General, who may seek civil penalties of up to 1 million dollars for a first violation and up to 3 million dollars for any subsequent violation. These figures are notable: the originally signed statute set them at 10 and 30 million dollars, and the March 2026 chapter amendments cut them by an order of magnitude — part of a broader alignment with California, whose own law caps penalties at 1 million dollars per violation without tiering. Day-to-day administration falls to a new oversight office within the Department of Financial Services, an unusual choice that places frontier-AI supervision alongside New York's banking and insurance regulator.

The federal shadow

The RAISE Act, like SB 53, arrives precisely as the federal government moves to contain laws of this kind. Executive Order 14365 of 11 December 2025 directs the Attorney General to stand up a litigation task force — formally established by memorandum on 9 January 2026 — whose sole task is to challenge state AI laws thought to burden interstate commerce or to be preempted. As of this writing the task force has not sued over any state statute, and an executive order cannot by itself displace state law; preemption would require an Act of Congress, which has not passed. For now the RAISE Act stands, and its 1 January 2027 start date gives developers a fixed horizon.

Why this matters beyond New York

For a European or other non-US organisation, the practical reading mirrors the California case. The frontier laboratories on which most downstream AI deployment ultimately depends will now publish — in two American states — the safety frameworks and incident histories of their largest models, a transparency floor that did not exist a year ago and that feeds directly into supplier due diligence. And the convergence is the real signal: New York and California, legislating independently, settled on the same compute and revenue thresholds. As our analysis of international AI governance notes, jurisdictions that disagree on much of their AI law are converging on a shared yardstick for the most powerful models — even as the exact enforcement design, from penalty tiers to which agency holds the pen, still differs.