Enki

New York opens AI lab registration under RAISE Act

By

Published

Reporting from TechCrunch

From January 1, 2027, New York's RAISE Act makes frontier AI labs with over $500 million in revenue register, publish safety plans and report incidents.

What it means for founders

  • Most startups sit outside it. The heavy duties need frontier scale compute and over $500 million in revenue, and a company calling a model by API is not the developer the law names.
  • Fine-tuning is the edge case. Since post-training compute counts, Cooley advises firms planning to "fine-tune, retrain or otherwise materially modify frontier models" to check whether they cross the line.
  • Better vendor paperwork. Transparency reports and big lab frameworks become public records worth tracking as platform risk.
  • Two clocks. Cooley notes California's SB 53 allows 15 days to report an incident, not New York's 72 hours, and caps penalties at $1 million per violation. New York also dropped its frontier specific whistleblower section.

Watch DIGIT's rules, which may add reporting duties, and the public list of registered developers it must keep.

What we don't know yet

The department sets each company's share of costs, assessed pro rata. A state spokesperson declined to say how enforcement against companies that skip registration would work, per TechCrunch. DIGIT's first annual incident report is due in 2028, and individual reports are exempt from public records requests. Related: Anthropic's draft bill on halting risky AI and policy coverage.

The story

New York opened a registration portal on Tuesday for the largest frontier AI developers, TechCrunch reports; companies such as OpenAI and Anthropic must sign up by January 1, 2027, when the RAISE Act takes effect. Governor Kathy Hochul signed the amended text on March 27, law firm Cooley says.

Who the law covers

Under the amended RAISE Act, a frontier developer is anyone who trained a foundation model using more than 10 to the 26th power operations, counting later fine-tuning and reinforcement learning on an earlier model. A large frontier developer is one whose revenue, including affiliates, topped $500 million in the previous calendar year.

It reaches models developed, deployed or operating at least partly in New York. Academic AI research at accredited New York colleges and universities, and the Empire AI consortium, are exempt.

What it requires

Every frontier developer must post a transparency report for each new or substantially changed model, covering release date, languages, intended uses and restrictions. A model card counts. Critical safety incidents, like stolen weights causing injury or a model deceiving its developer to evade its controls or monitoring, outside a test built to provoke that, in a way that shows materially higher catastrophic risk, go to the state within 72 hours. One threatening imminent death or serious injury must reach a relevant authority within 24 hours.

Large frontier developers must also publish a frontier AI framework on risk thresholds, mitigations, third party testing and weight security, review it yearly, and post material changes within 30 days. They send summaries of their catastrophic risk assessments from internal model use every three months, or on a schedule the office agrees to, help pay the regulator's running costs, and file an ownership and contact disclosure, renewing it at least every two years.

Who enforces it

Oversight goes to the Office of Digital Innovation, Governance, Integrity and Trust (DIGIT) inside the Department of Financial Services, the governor's office said in September. The attorney general can sue a large frontier developer for civil penalties of up to $1 million for a first violation or $3 million for each later one. After notice and a hearing, DIGIT can also impose $1,000 per day that a disclosure goes unfiled or false information uncorrected. There is no private right of action.

Sources

Primary sources

Reporting

Tools in this story

We may earn a commission if you sign up through our links. It never affects our ratings or which stories we cover.

Claude

Anthropic's assistant for writing, analysis, and code

9.0Editor’s scoreVisit Claude
ChatGPT

OpenAI's all-purpose AI assistant

9.2Editor’s scoreVisit ChatGPT

More in Policy & Safety

How Enki covers newsCorrectionsReport an error