What did New York open on 10 November 2026?
New York opened the RAISE Portal on 10 November 2026, giving large frontier AI developers a state registration window that closes on 1 January 2027. The portal is run by a new office called DIGIT, short for Digital Innovation, Governance, Integrity, and Trust.
DIGIT sits inside the Department of Financial Services. The governor's office said the portal lets large frontier developers register and begin complying with the RAISE Act's safety, transparency and reporting duties. The office also published a website describing what developers must do.
Governor Kathy Hochul paired the launch with a new research center at SUNY Binghamton. The New York Center for AI Responsibility and Research will bring together more than 80 researchers to test advanced AI systems. It will also help review the safety frameworks that companies publish.
The law itself is not new. Hochul signed the RAISE Act on 19 December 2025, then signed a chapter amendment that finalized it on 27 March 2026, according to law firm Wiley. The substantive duties begin on 1 January 2027.
“The stakes are too high to simply take AI companies at their word,” Hochul said in the announcement. Her office described DIGIT as the first office of its kind in the country.
The launch puts New York alongside California in building an operating frontier-AI regime. “While the federal government fails to act, New York is leading the way on AI safety and holding big tech accountable,” Hochul said.
TechCrunch framed the day as oversight arriving for OpenAI and Anthropic. The statute does not name companies. It uses revenue and computing thresholds instead, which is why the scope question matters so much for labs of different sizes.
Who has to register under the New York RAISE Act?
The statute separates frontier developers from large frontier developers. The split decides who registers, who publishes a safety framework and who pays for the state's oversight.
A frontier model is a foundation model trained with more than 10^26 integer or floating-point operations, according to the bill text. The count includes the original training run and any later fine-tuning, reinforcement learning or material modification.
A foundation model must be trained on a broad data set, designed for generality of output and adaptable to many tasks. A frontier developer is anyone who trained, or began training, a model that crosses the compute line.
A large frontier developer is a frontier developer that, together with affiliates, had annual gross revenue above $500 million in the prior calendar year. Only this larger group must file the disclosure statement by 1 January 2027.
The compute threshold separates frontier training runs from smaller models without naming a parameter count or a price. Coverage therefore depends on information that is often private, including how much compute a company used.
The scope clause reaches beyond New York companies. The law applies to frontier models “developed, deployed, or operating in whole or in part” in the state. A model built in California and served to New York users can fall inside it.
“Deploy” has a defined meaning too: making a model available to a third party for use, modification, copying or combination with other software. Sharing a model only for development or evaluation does not count.
A company with high revenue but no model above the compute line is not a frontier developer. A company with a qualifying model but lower revenue still has incident and transparency duties, even though it does not register.
Two groups are exempt. Accredited colleges and universities are exempt for academic research, and so are the Empire AI consortium and its institute. Everything else above the thresholds is covered.
What must covered AI developers publish and report?
Large frontier developers must publish a frontier AI framework on their websites. The document must explain how the company identifies catastrophic risk, applies mitigations and secures unreleased model weights.
The bill lists ten required topics. They include national and international standards, risk thresholds, third-party evaluations, annual updates, cybersecurity, internal governance and incident response.
The framework must also cover assessments of catastrophic risk from internal use, including the risk that a model circumvents oversight. Companies must review the framework at least once a year, and they have 30 days to publish and justify a material revision.
Transparency reports are a separate duty. Before or when a developer deploys a new frontier model, it must publish the release date, supported languages, output types and intended uses.
The report must also summarize catastrophic-risk assessments and say how far outside evaluators were involved. A company can satisfy the rule by putting the material in a model card or system card.
Redactions are allowed for trade secrets, cybersecurity, public safety and national security. Each redaction must be described and justified, and the unredacted material must be kept for five years.
Incident reporting carries the tightest deadline. A frontier developer must report a critical safety incident to DIGIT within 72 hours of determining that one occurred.
The clock also runs from the moment the developer learns facts that create a reasonable belief an incident happened. If the incident poses an imminent risk of death or serious physical injury, a report must also reach law enforcement or a public safety agency within 24 hours.
Large developers send DIGIT quarterly summaries of catastrophic-risk assessments from internal use. The office may agree to a different schedule if the company requests one in writing.
Registration is more than a name and an address. The disclosure statement must identify the company, its New York offices, its contacts and, for private firms, owners of 5% or more.
It must be renewed every two years, or sooner after an ownership change or a material change in the information. Large developers also pay a pro rata share of the office's operating costs.
Missing or false filings can trigger a $1,000-per-day penalty after notice and a hearing. The office can also recover unpaid assessments.
The public has a channel as well. The office must set up a mechanism for anyone to report a suspected critical safety incident, and it may review those reports alongside company filings.
Incident reports and quarterly risk summaries are exempt from disclosure under New York's public officers law. The office may share them with other agencies, including the attorney general, at its discretion. That design gives regulators more information than the public will get.
How will New York enforce the RAISE Act?
The attorney general enforces the law through civil actions. Penalties reach $1 million for a first violation and $3 million for each later violation, scaled by severity.
Covered failures include not publishing a required document, not reporting an incident, making a false or misleading statement, or ignoring the company's own published framework.
There is no private right of action. People and advocacy groups cannot sue under this statute, though other laws may still apply.
The office has a separate lever for registration. It can impose the daily penalty, recover assessments and publish a list of large developers that have filed.
The law does not create a pre-approval regime. Its main tools are disclosure, registration and civil penalties applied after a violation.
From January 2028, the office must produce an annual report with anonymized and aggregated incident information. The report can recommend changes to the law and must avoid material that would expose trade secrets or national security.
The law also lets the office recognize federal reporting standards as an equivalent route. A developer that takes that path must declare its intent and still copy federal incident reports to the state.
The office can adopt a federal standard only if it is substantially equivalent to, or stricter than, New York's reporting rule. The designation can be revoked if the federal rule stops meeting that test.
How does New York compare with California and Illinois?
California's Transparency in Frontier Artificial Intelligence Act, SB 53, took effect in January 2026. It requires published safety frameworks, whistleblower protections and incident reports within 15 days.
New York's 72-hour window is far shorter, and its framework duties land on 1 January 2027. Wiley calls the incident deadline a marked departure from California's approach.
Illinois passed its Artificial Intelligence Safety Measures Act, known as SB 315. It also takes effect on 1 January 2027 and uses the same 10^26-operation and $500 million thresholds, Pillsbury says.
Illinois adds independent safety compliance audits, according to TechCrunch. It also allows penalties of up to $1 million for a first violation and $3 million after that.
Pillsbury says Illinois will require publication and implementation of a frontier AI framework from 2028, a year after its first disclosure filings.
Connecticut's Public Act 26-15 concentrates on whistleblowers. It requires anonymous internal reporting processes at large frontier developers by 1 January 2027.
Companies that operate in several states will need one incident process that can meet the shortest deadline. They will also need frameworks that satisfy different publication rules. That is the compliance cost the laws impose even where their substance overlaps.
The patchwork creates a political problem as well. Wiley notes that the New York law is “in tension with the Trump Administration's efforts to promote comprehensive federal AI legislation.”
What does the new research center add?
The New York Center for AI Responsibility and Research is based at SUNY Binghamton. It will convene more than 80 scientists from colleges and research institutions across the state.
Greg Morrisett, dean of Cornell Tech, and Jeannette Wing, Columbia's executive vice president for research, are interim co-directors. Both have worked on how complex computing systems fail and how to secure them.
The center will use Empire AI, New York's $500 million public-private computing initiative. Its work covers independent testing, foundational research and training new safety experts.
A $30 million philanthropic gift led by Tom Secunda, a Binghamton alumnus and Bloomberg LP co-founder, helped launch it. SUNY added $25 million for research, and the state provides operating capital.
The center can advise but not regulate. At the state's request, it can assess systems and recommend changes, and it will coordinate with DIGIT when reviewing published frameworks.
DFS acting superintendent Kaitlin Asrow said the department is proud to lead implementation through DIGIT. The governor's office said the center can provide expertise and evidence to inform state oversight.
Its first recruiting push targets formal methods, hardware security, interpretability, alignment, agentic systems and biosecurity. That mix points to where New York expects the next safety questions to come from.
What are the open questions and what happens next?
Registration is the first test. Each lab must decide whether its models cross the compute threshold and whether its revenue, including affiliates, crosses $500 million.
The second test is disclosure quality. Redactions are allowed for trade secrets, cybersecurity, public safety and national security. Each redaction must be justified, but the public will see less than regulators do.
The definitions are narrow in places. A critical safety incident generally requires death, injury or realized catastrophic harm. Loss of control and weight theft count only when they cause death or bodily injury, and deceptive behavior counts when it shows materially increased risk.
Catastrophic risk has its own bar: a foreseeable and material risk of death or serious injury to more than 50 people, or more than $1 billion in property damage, from a single incident. Many harmful outputs will not meet either threshold.
Enforcement capacity is another open question. DIGIT is new, and Pillsbury notes that the September announcement did not explain the form or substance of required submissions.
Staffing will matter as much as rules. Pillsbury reported that DFS personnel with technology, cybersecurity and AI expertise were supporting the office, with more dedicated staff expected.
Companies near the thresholds face their own judgment calls. They must decide whether affiliates or later fine-tuning cross the line, and the answer may change as models are updated.
DIGIT also has rulemaking authority. It can add reporting or publication requirements, including post-incident information, if it decides the additions help safety and transparency. That power could expand the paperwork before the first compliance cycle ends.
Federal preemption is the biggest wildcard. If Congress passes a national framework that displaces state rules, the portal could become a compliance layer rather than the final word.
OpenAI has said it supports the RAISE Act and wants mandatory national safety requirements from Congress. That position gives New York's experiment an unusual ally among the labs it regulates.
The dates to watch are 1 January 2027 for registration and first duties, then January 2028 for the office's first annual report. The published list of registrants will show how many labs concluded they are covered.
