Skip an AI Safety Filing in New York and the Fines Start at $1 Million

New York has opened registration for large frontier AI developers, starting a compliance process that will make safety disclosures legally required and expose failures to million-dollar penalties. ^[2][3]

The portal opened on November 10, 2026. Covered large developers must register by January 1, 2027, with the RAISE Act's main compliance requirements beginning in January. Those include publishing safety frameworks, reporting critical incidents and submitting regular assessments of catastrophic risks. ^[2][6]

Companies including OpenAI and Anthropic face the new oversight, according to TechCrunch. The law targets developers of powerful frontier models, rather than every business that uses an AI tool, and places additional obligations on developers above its revenue threshold. ^[1][2][4]

The change gives New York a way to pursue failures that could previously have amounted to breaking a company's voluntary commitments. The attorney general can bring civil actions against large frontier developers for specified violations, including missing required reports or making false statements. Penalties can reach one million dollars for a first violation and $3 million for subsequent violations. ^[3]

That is a meaningful change in the cost of ignoring a disclosure requirement. Whether it produces better safety will depend on the state's ability to check what developers submit and identify what they leave out.

Who must comply

RAISE stands for the Responsible AI Safety and Education Act. Governor Kathy Hochul signed it in December 2025, and its requirements were subsequently shaped through amendments. The state's rationale was that developers already had voluntary safety frameworks, but their disclosures were inconsistent and insufficiently standardized. ^[3][4][6]

New York describes two levels of responsibility. All frontier developers operating in the state will face model-transparency and critical-incident reporting requirements. Large frontier developers have extra duties, including registration and published safety and transparency frameworks. ^[2][6]

The Assembly's amendment bill text defines a frontier model by the computing work used to train it, with a threshold above 10^26 operations. It defines a large frontier developer as one whose annual gross revenue, together with that of its affiliates, exceeded $500 million in the preceding calendar year. These definitions appear in the legislative text, rather than in a published list of covered companies. ^[4]

The revenue test therefore looks beyond an individual developer's own accounts to include affiliated businesses. The model test looks at training resources. Both matter to coverage, so the law's reach cannot be reduced to a list of familiar chatbot brands. ^[4]

New York has not yet established publicly which companies DIGIT will identify as covered registrants, how it will interpret the incident-reporting clock or what verification will precede enforcement. No penalty under the new compliance regime has been reported so far. ^[1][2][6]

A safety framework becomes a filing obligation

Large developers must establish and publish safety and transparency frameworks, register with the state, pay assessments and file quarterly catastrophic-risk assessments. They must also file a disclosure statement at least every other year, starting January 1, 2027. ^[6]

Catastrophic risk means the potential for large-scale harm, such as mass casualties or major damage. ^[4]

Those duties create several different checks on a developer. A published framework sets out its approach for public scrutiny. A quarterly assessment requires it to revisit catastrophic risks regularly. Registration and recurring disclosures give the state a continuing relationship with the company, rather than a single announcement of good intentions. ^[6]

Critical safety incidents bring a faster deadline: large developers must report them to the state's Office of Digital Innovation, Governance, Integrity and Trust, known as DIGIT, within 72 hours. DIGIT sits within New York's Department of Financial Services and is responsible for implementing the law. ^[2][6]

The incident category concerns serious harm. The Assembly text includes harm arising from catastrophic risks, loss of control that causes death or bodily injury, and certain unauthorized access to model weights that results in death or bodily injury. Model weights are the numerical settings learned during training that govern how a model responds. ^[4]

A disappointing chatbot answer is a long way from those examples. The reporting system is directed at events with grave consequences, which makes the state's interpretation of the threshold important for both developers and the public.

Developers will not be the only route by which incidents reach the state. Members of the public may submit suspected critical safety incident reports to DIGIT. The office may pass reports to other government authorities and is to publish an annual public report. ^[6]

That gives the regulator a potential check against company submissions. A public report is still a report of a suspected incident, however, not an established violation.

The test moves from promises to oversight

New York is also building technical capacity alongside the filing system. Hochul said a new state research center at Binghamton University would independently test powerful AI systems. The center is intended to work with DIGIT to review developers' published safety frameworks and make recommendations. It is not itself the enforcement authority. ^[2]

The split matters in practice: technical reviewers can examine a safety framework, while the attorney general has the authority to pursue the specified civil violations. A fine requires an enforcement process, not merely an unfavorable assessment. ^[2][3]

Anthropic's policy page says the company has advocated for safety-transparency laws, including RAISE. It also says transparency alone is insufficient to address the most serious risks. ^[5]

That is the right limit to put on this law's promise. Disclosure creates material that can be checked and duties that can be enforced. It does not, by itself, demonstrate that a model is safe.