Skip to main content

New York RAISE Act Explained: Who It Covers and When It Starts

Title card for the Vaquill AI guide: New York RAISE Act Explained: Who It Covers and When It Starts

Short answer: no, the New York RAISE Act is not in force yet. It takes effect on January 1, 2027. The Responsible AI Safety and Education (RAISE) Act covers frontier developers, which are companies that train the very largest AI models. The governor signed it on December 19, 2025. A chapter amendment, which is a follow-up law that edits one already signed, then rewrote it. The governor signed that on March 27, 2026. The state's own code text and announcements treat the amended version, starting January 1, 2027, as the law.

The heaviest duties fall on "large frontier developers," those with over $500 million in yearly revenue. The attorney general (AG) enforces it, with fines up to $1 million for a first violation and $3 million for each later one. This is not legal advice. Check the current text before you rely on it.

Reviewed and corrected on 12 October 2026. This is general information, not legal advice; check the cited statute or opinion before relying on it.

Official text, as of 12 October 2026

New York General Business Law § 1420(9)(a), added by Chapter 96 of the Laws of 2026: a "frontier model" is "a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations." Section 1427(2) adds: "Nothing in this article shall be construed to establish, authorize or create a private right of action associated with violations of this article." The Senate's page for section 1427 carries the note "Effective January 1, 2027." The signed chapter amendment is S8828.

What changed in 2026: Chapter 96 repealed the article as first signed in December 2025 and added a new Article 44-B, with a $500 million revenue test for "large" developers, lower fines, and a January 1, 2027 start.

TL;DR

  • Status on October 12, 2026: signed, amended, and not yet effective. The official start is January 1, 2027. On September 21, 2026, New York said large developers will be directed to register starting in November.
  • The bill did not change between passage and signing. It was signed as passed. A chapter amendment then changed it. Guides that describe the December version describe text the state no longer treats as the law.
  • Who: developers who trained a model with more than 10^26 operations (a count of basic computer arithmetic steps). Large ones, over $500 million in revenue, carry the penalty.
  • Enforcement: the AG sues. A new office in the Department of Financial Services (DFS) takes reports, and the department bills large developers for its costs. The text creates no private right to sue.
  • California's Senate Bill (SB) 53 is close in design but started a year earlier. A table below compares them.

Why you see two different versions

Most coverage of this law dates from December 2025 or January 2026. That is before the chapter amendment became law. Some of it still describes the signed version. That version had a test based on $100 million spent on computing power, and fines up to $10 million. The final version has a $500 million revenue test and fines up to $3 million. A search for the "New York RAISE Act" mixes the two.

A statute is a law passed by a legislature. A session law is the text of a bill as signed. The code is the organized set of statutes. Here the code is the General Business Law (GBL), Article 44-B, sections 1420 to 1429. For how the two relate, see what a session law is.

New York RAISE Act effective date and key dates

The dates come from the Assembly's bill records, Chapter 699 of 2025, Chapter 96 of 2026, and the DFS announcement of September 21, 2026.

  • June 12, 2025: the Senate and Assembly passed S6953-B and A6453-B. Both houses passed it that day, according to the Assembly's bill record.
  • December 19, 2025: the governor signed it as Chapter 699 of the Laws of 2025. The signed text said it would take effect "on the ninetieth day after it shall have become a law."
  • January 6 and 8, 2026: the chapter amendment was introduced in the Assembly as A9449 (January 6) and in the Senate as S8828 (January 8).
  • January 28, 2026: the Senate passed S8828.
  • February 25, 2026: the Assembly substituted S8828 for A9449. S8828 is the bill that became law.
  • March 11, 2026: the Assembly passed S8828. It was delivered to the governor on March 20.
  • March 27, 2026: signed as Chapter 96 of the Laws of 2026. Section 3 of the amendment swaps the signed start clause for "January 1, 2027." The old clause read "on the ninetieth day after it shall have become a law."
  • September 21, 2026: the governor and DFS announced implementation steps. "Starting in November, New York will direct large frontier AI developers to register with the State." They also named the new office inside DFS. It is the Office of Digital Innovation, Governance, Integrity and Trust (DIGIT).
  • January 1, 2027: the law takes effect. Large developers must file a disclosure statement, "at least every other year, starting January 1, 2027," says the release.
  • January 1, 2028: the DFS office's first annual public report is due (GBL § 1422(7)).

The amendment replaced the whole article. Section 2 of Chapter 96 says the article "is REPEALED and a new article 44-B is added." The amended article is the one in the state's code.

One timing point needs care. The December text had a 90-day start, which would have fallen on March 19, 2026. The amendment passed the Legislature on March 11, reached the governor on March 20 and was signed on March 27. The Senate's code text for the article's sections carries the note "Effective January 1, 2027," and the September 2026 release says "Beginning January 2027." In our October 2026 review we found no source that treats any earlier date as the start. If the gap between those dates matters to you, ask a lawyer.

What the amendment changed

We compared both versions on the Assembly's and Senate's sites. Here is what moved.

  • Who is "large." As signed: a developer that had spent over $100 million on computing power. As amended: yearly revenue over $500 million.
  • Model line. As signed: over 10^26 operations and over $100 million in computing cost. As amended: over 10^26 operations.
  • Safety document. As signed: a "safety and security protocol." As amended: a "frontier AI framework."
  • Ban on release. As signed: a large developer could not deploy a model if that "would create an unreasonable risk of critical harm." As amended: no such ban.
  • Harm measure. As signed: death or serious injury of 100 or more people, or at least $1 billion in damages. As amended: more than 50 people, or more than $1 billion in damage.
  • Incident reports. As signed: to the AG and a state security agency, in 72 hours. As amended: to a DFS office, in 72 hours.
  • Top fine. As signed: $10 million, then $30 million. As amended: $1 million, then $3 million.
  • Start. As signed: 90 days after signing. As amended: January 1, 2027.

Who it covers

Section 1420(9)(a) defines a "frontier model" as "a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations," meaning a very large count of basic calculations done during training. A foundation model is a general-purpose AI trained on broad data. The compute count includes the original training run and any later fine-tuning or reinforcement learning the developer applies to a preceding foundation model (section 1420(9)(b)). A "frontier developer" is a person who trained, or began training, such a model.

Section 1420(10) defines a "large frontier developer." Its yearly revenue, with its affiliates (related companies under common control), must be "in excess of five hundred million dollars" in the preceding year. Section 1425 limits the article to frontier models "developed, deployed, or operating in whole or in part in New York state." Section 1426 exempts colleges doing academic research. It also exempts the Empire AI consortium.

The 10^26 line is extremely high and aimed at frontier model builders. If you only use someone else's model, these duties sit with the builder. A startup that builds an app on a vendor's model is a user. The lab that trained the model is the developer.

What it requires

  • A published safety framework. A large frontier developer must publish a "frontier AI framework." It must "describe in detail" how the developer handles risk testing, safeguards, security and incident response. The developer reviews it every year (§ 1421).
  • A report with each release. Any frontier developer must post a transparency report before or with a new model. Large ones add summaries of their risk checks (§ 1421(3)).
  • Incident reports in 72 hours. A "critical safety incident" (section 1420(4)) is one of four things: unauthorized access to a model's weights that results in death or bodily injury; harm from a catastrophic risk coming true; loss of control of a model causing death or bodily injury; or a model using deceptive techniques against its developer to subvert its controls, outside an evaluation designed to elicit that behavior, in a way that shows materially increased catastrophic risk. A frontier developer must report one to the DFS office "within seventy-two hours" (§ 1422(3)). The clock starts when it determines the incident happened, or learns facts that give a reasonable belief. If an incident poses an imminent risk of death or serious injury, the developer must tell the right authority within 24 hours.
  • Registration and fees. A large frontier developer needs a current disclosure statement on file with the office. It must also pay its share of the office's costs (§ 1428). The statement is renewed every two years, or sooner if ownership of the model is transferred or a material fact changes.
  • Quarterly risk summaries. A large frontier developer sends the office a summary of its assessments of catastrophic risk from internal use of its models every three months, or on another reasonable schedule the office agrees to (§ 1422(2)). The DFS release lists this among the duties.
  • No false statements. A frontier developer may not make a materially false or misleading statement about catastrophic risk from its models or its management of that risk, and a large one may not misstate its implementation of its own framework (§ 1421(4)(a)). A statement made in good faith and reasonable under the circumstances is outside the rule (§ 1421(4)(b)). A catastrophic risk means a foreseeable and material risk of death or serious injury to more than 50 people, or more than $1 billion in damage, from a single incident in which a frontier model gives expert-level help with a chemical, biological, radiological or nuclear weapon, acts with no meaningful human oversight in a cyberattack or in conduct that would be murder, assault, extortion or theft if a human did it, or evades the control of its developer or user (§ 1420(3)).

New York RAISE Act penalties and enforcement

Section 1427(1) lets the attorney general "bring a civil action to recover a civil penalty." A civil action is a lawsuit in court. A civil penalty is a fine paid to the state. The cap is "one million dollars for a first violation" and "three million dollars per subsequent violation." It applies when a large frontier developer fails to publish a required document, makes a false statement, fails to report an incident, or breaks its own framework.

Section 1428(5) adds a separate penalty run by the DFS office. After notice and a hearing, the office may levy "one thousand dollars for each day" a large frontier developer operates without a current disclosure statement or fails to correct false information in it, plus an amount equal to unpaid assessments.

On private lawsuits the text is direct. Section 1427(2) says: "Nothing in this article shall be construed to establish, authorize or create a private right of action associated with violations of this article." A private right of action is a right for a private person or company to sue under the law. The article has no whistleblower section.

Section 1429 lets the office write rules. The September 2026 release says the governor is also "exploring ways to build on the RAISE Act." The article can change again.

How it compares with California's SB 53

The two laws share definitions almost word for word. They differ in speed, reporting and who gets the reports. California's SB 53 is Chapter 138 of the Statutes of 2025, approved September 29, 2025. It sits in Business and Professions Code sections 22757.10 to 22757.16, and the table follows that signed text. For the California side, see California AI laws explained.

PointNew York (GBL Art. 44-B)California (SB 53)
StartsJanuary 1, 2027January 1, 2026
Model lineOver 10^26 operationsSame
"Large" lineRevenue over $500 millionSame
Incident report72 hours15 days
Imminent danger24 hours24 hours
Gets the reportsAn office in DFSOffice of Emergency Services
Top fine$1 million first, $3 million later$1 million per violation
Who suesAttorney generalAttorney general
Registration and feesYes, every two yearsNone
Whistleblower rulesNone in the articleLabor Code § 1107.1
First public reportJanuary 1, 2028 (§ 1422(7))January 1, 2027 (§ 22757.13(g))

What a business should check

  1. Find out whether you trained a model above the 10^26 line, or plan to. If not, you are a user.
  2. If you use a vendor's model, ask which of its duties under this law and SB 53 it says it meets. Ask for the framework link. A startup that builds on a vendor's model usually does not file anything itself. The lab that trained the model does.
  3. If you are a large frontier developer, watch for the DIGIT registration notice that DFS said would start in November 2026.
  4. Put January 1, 2027 on your calendar for the first disclosure statement.
  5. Check the Assembly or Senate bill page and DFS every quarter. Rules can still come.

What goes wrong

The common mistake is citing the December 2025 version. Its $100 million test, 100-person harm line and $10 million fine are gone. The next is treating the signing date as the start date. The law starts January 1, 2027. The last is assuming a user of a big model is covered. The article reaches builders at the frontier.

For another state's approach, read the Colorado AI law guide. To cite New York sections correctly, see how to find and cite New York law.

Where to read the official text

The Assembly's and Senate's pages show each version and its history. The September 21, 2026 facts above come from the DFS announcement of that date. Here are the pages: A6453-B, the bill as signed in December 2025, S8828, the chapter amendment that became Chapter 96, and its Assembly companion A9449, for which S8828 was substituted. The DFS announcement of September 21, 2026 covers registration. The current code text is on the New York Senate's laws site under General Business Law sections 1420 to 1429.

This guide is part of US Law Data: The Complete Guide, a map of where US law comes from and how to use it.

Spot the problem
Question 1 of 4

A 2025 blog post says New York's RAISE Act fines large AI developers up to $10 million and bans releasing a model that creates an unreasonable risk of critical harm. Your GC asks if that is still the law. What do you tell her?

FAQ

Is the New York RAISE Act in effect? Not yet. It was signed and then amended. The state's code text and its September 2026 release put the start at January 1, 2027.

What is the RAISE Act effective date? January 1, 2027. The first signed text said 90 days after signing. The March 27, 2026 amendment replaced that with the January date.

Who does the RAISE Act apply to? Frontier developers, meaning companies that trained a model with more than 10^26 operations. The fines and the registration duty fall on large frontier developers with over $500 million in revenue. If you only use another company's model, the main duties sit with the builder.

What are the penalties? The attorney general can seek up to $1 million for a first violation and up to $3 million for each later one. The DFS office can add $1,000 per day for a missed disclosure filing.

Can someone sue me under it? The text says nothing in the article creates a private right of action. Only the attorney general sues for the penalty.

How is it different from California's SB 53? New York starts a year later. It gives 72 hours to report an incident, where California gives 15 days. It adds registration and fees. It has no whistleblower section.

Does it apply if I just use an AI model? Usually the main duties sit with the company that trained the model, not the business that only uses it.

For the US statutes, regulations and court rules these laws sit in, see the US primary law API.

Connect our US primary law database.
Every US statute, regulation, constitution, and executive order via REST, MCP or SQL. 6M+ sections, section-level citations, and links to the official source. Plus a free open dataset.
Updated October 12, 202618 min read

New legal AI guides, weekly.

Priyansh Khodiyar

Priyansh Khodiyar

Co-Founder & CTO

Priyansh leads engineering and AI at Vaquill AI: the pipelines that pull statutes, regulations and court rules from every US jurisdiction's official publisher, and the REST API, MCP server and open dataset that serve them.

Cookie preferences

Choose what Vaquill may use on this browser. You can change this at any time from “Cookie settings” in the footer.

Details in the Cookie Policy.