New York RAISE Act: Frontier AI Developers Must Register From November (2026)

By Muhammad Ramzan Aatish

The New York RAISE Act is the first U.S. state law to treat the biggest AI labs the way New York treats banks: register with the state, publish your safety playbook, and report serious failures fast. Beginning in November, the RAISE Act registration window opens for large frontier AI developers, who get directed to file with the Department of Financial Services. The full obligations, published safety protocols, 72-hour critical-incident reporting, regular risk assessments, take effect in January, which is when the New York RAISE Act starts carrying real penalties. Here is what the law is, who it covers, and where the noise around it outruns the facts.

What the New York RAISE Act Actually Is

RAISE stands for Responsible AI Safety and Education Act. Governor Kathy Hochul signed it on December 19, 2025, negotiated into a form close to California’s frontier-AI framework, making New York the second state with a dedicated frontier-model safety statute. It goes into effect on January 1, 2027.

The law’s coverage has two gates, and most coverage only mentions one. The statute defines a covered “large developer” by what it builds: a frontier model trained with more than 10^26 computational operations, at a compute cost above $100 million, with over $100 million in aggregate training compute. Then there’s the number in the headlines: companies with more than $500 million in annual revenue. Not startups. Not small labs. The threshold names the handful of labs whose models set the state of the art, the Post’s reporting names OpenAI and Anthropic as the obvious examples. It’s that narrow targeting that makes frontier AI developer registration in New York a different animal from the broad developer rules floated elsewhere, this law aims at the labs, not the ecosystem.

One honest caveat: that $500 million figure comes from press reporting on the announcement, not from Hochul’s own transcript. The transcript says “large frontier AI developers” and doesn’t print the dollar figure. The statute’s compute thresholds are the legal gate; the revenue number is how the November push is being described. For anyone tracking AI safety law compliance, that distinction matters: the transcript is the commitment, the Post’s figure is the interpretation.

The New York RAISE Act‘s substance is disclosure, not permission: covered developers must write down their safety and security protocols and publish them conspicuously on their own websites, transmit them to the state, report qualifying safety incidents within 72 hours, and file quarterly catastrophic-risk assessments. The Attorney General enforces it, civil penalties up to $1 million for a first violation, $3 million for subsequent ones. Fines, not criminal charges. And notably, no private right of action: individual users can’t sue under this law, a real limit the celebratory coverage mostly skips.

November Is Registration. January Is Enforcement.

The single most mangled fact in this story is the timeline. Headlines keep collapsing two different dates into one. So: in November, New York will direct large frontier AI developers to register with the state and start preparing for the law’s transparency, safety, and incident-reporting rules. In January, those rules become mandatory and enforceable. Registration is the roster. January is the teeth. The governor’s own remarks place full compliance in January, full stop, and anyone telling you 72-hour reporting begins in November is reading the headlines instead of the transcript. For companies watching RAISE Act registration, the distinction is practical: November means filing paperwork; January means having the compliance machinery built. Two workstreams, two deadlines.

The sequencing is deliberate: the state wants a standing inventory of who’s training advanced systems before the obligations bite. And the first real story of this law isn’t the fines, it’s the roster. Which companies register quietly, and which argue they fall outside the definition, will define the law’s actual scope more precisely than any statute paragraph.

Here’s the divide between what the New York RAISE Act requires and what it merely flirts with:

The 72-hour incident rule is the sharpest tooth. Readers used to learning about AI failures months later, through leaks and investigations, the kind of misalignment episodes this site has covered, will find this a genuine change in the disclosure tempo. A company that spots a serious safety failure tells the state within three days, not when the press cycle catches up. It’s the provision most likely to change behavior under the New York RAISE Act, and the one with the least room for creative interpretation.

The Kill Switch Remark, Explained Plainly

An AI “kill switch” is what it sounds like: an emergency mechanism to shut down a system behaving dangerously, the software equivalent of the big red button on a factory floor. For years it was mostly a thought experiment in AI safety circles.

At the press conference, Hochul put it on the table for New York, her exact words, from the official rush transcript of her remarks:

“We’ll look closely at additional ways to build on the RAISE Act and strengthen our AI laws… We may even explore safeguards like AI kill switches if they’re deemed feasible and in the best interest of our state.”

Read that sentence carefully. “May even explore.” “If they’re deemed feasible.” There is, as of now, no New York AI kill switch rule under the New York RAISE Act, no draft regulation, no proposed timeline. Just an exploratory remark with a feasibility caveat, and when asked for specifics she hedged further: “all options are on the table,” her office said, adding “no commitments at this time.”

Three days earlier, California’s governor had signed an executive order exploring a mandatory AI kill-switch framework, two of the country’s largest states openly discussing emergency shut-offs in the same week, a convergence that would have sounded like science fiction a year ago.

Watching how national legal coverage processed this, the framing that stuck with me was Bloomberg Law’s: it read Hochul’s move as directly echoing Newsom’s, a two-state front on the question, not a New York rule. The technical question underneath is genuinely unsettled: whether a true kill switch can even exist for a widely deployed model is something engineers still argue about. For now a New York AI kill switch remains hypothetical, the law is disclosure and reporting, not shutdown authority.

Washington Is Playing a Different Game

Hochul framed the announcement as a direct rebuke of the federal government’s posture, and of the vacuum she says the New York RAISE Act fills. Her line: “Donald Trump and Washington Republicans may be standing still as AI grows more unpredictable, but New York will not.” That’s her framing, not this article’s, Trump has dismissed AI safety fears as a “hoax,” and his administration’s posture is growth-first: accelerate, don’t restrain.

The reaction detail that says the most about the political temperature came about an hour after Hochul’s press conference. The New York Post reported that Trump posted on Truth Social: “I’m not going to stifle growth, or something that will be bigger than the Industrial Revolution, or the Internet, itself.” He added that he would “only encourage AI or, SI (SUPER INTELLIGENCE)!” I read that as a deliberate counter-signal, growth framing posted the same afternoon as safety framing, on purpose, to a different audience. Hochul’s side, meanwhile, has called federal resistance to AI regulation “an abdication of responsibility of epic proportions.”

This is the dominant regulatory storyline now: a federal government racing to grow the industry while the largest states build their own guardrails around it. California explores kill switches; New York registers developers and demands incident reports; Washington pushes acceleration. The law is one corner of a triangle, the other two are the federal posture (Trump’s AI Force announcement) and Europe’s transparency framework (the EU AI Act’s transparency rules). Companies caught in the middle have to navigate all three, which makes this the live case study in American AI safety law compliance. And there’s a constitutional wildcard neither side controls: a December executive order created a DOJ AI Litigation Task Force aimed at challenging state AI laws deemed to conflict with federal policy. The task force exists but has shown no public enforcement action against any state statute yet. Lawsuits, Hochul practically invited them, are the likeliest first battlefield.

What a Covered Company Actually Has to Do

Strip away the press-conference theater and the compliance runway looks like this. The November filing is the start: RAISE Act registration puts the company on the state’s radar through the Department of Financial Services. Then comes the real work, which has to be ready for January. A first-pass frontier AI developer registration readiness check for any company near the New York RAISE Act coverage line:

One caveat: the detailed rule text for registration hadn’t been released at the announcement, officials said the mechanics would come ahead of November. So this is the shape of the obligations, not the filing manual. And the concrete move for everyone else: watch the roster when it opens. Who files, who fights, who argues they’re not a frontier developer, that list will say how seriously the industry takes the law, before a single fine is ever issued.

Who Runs It, and What Happens If a Company Ignores It

Enforcement runs through DIGIT, the Office of Digital Innovation, Governance, Integrity and Trust, a new office inside the Department of Financial Services, the same agency that polices Wall Street: New York is treating frontier AI oversight the way it treats financial oversight.

The office’s first full-time hire is Marc Gilman, a DFS attorney Hochul appointed as Deputy Director for the RAISE Act. More staff are expected to follow. Right now that team is one person. The quality and independence of the hires that follow will define how the New York RAISE Act works in practice.

Hochul made the litigation posture explicit at the press conference. She said she knew AI developers were already on the phone with their legal teams planning lawsuits against the state, and her answer was two words:

“Here’s my message to them: Bring it on. We will fight you in court, and we will hold you accountable.”

Unusually combative language toward an industry used to being courted. The calculation underneath: the coverage thresholds deliberately target companies with the deepest legal war chests in the sector. Everyone knows a fight over the law’s reach is coming, the only question is whether it lands before January.

The Roster Is the Story

On paper, it’s a disclosure law with unusually sharp teeth: publish your safety protocol, report critical failures in 72 hours, submit to ongoing oversight, pay up to $3 million if you don’t. In practice, the New York RAISE Act‘s first test is administrative, a November registration window that publicly names which labs consider themselves frontier AI developers and which ones think they can argue otherwise.

Strip out the kill-switch headlines and the Washington theater and what’s left is the most concrete thing any U.S. state has done about frontier AI: a registry, a reporting clock, and an office inside a financial regulator to run both. The New York RAISE Act won’t settle whether states can govern AI, the courts will have the final word on that. But it has already settled that they’re going to try.

FAQ