New York RAISE Act Updated: Aligning with California on Frontier AI Governance

New York has joined California at the frontier of AI regulation. With Governor Kathy Hochul signing the chapter amendments to the Responsible AI Safety and Education Act — commonly called the RAISE Act — in March 2026, New York finalized a governance framework for frontier AI models that tracks closely with California’s Transparency in Frontier Artificial Intelligence Act (TFAIA) enacted under SB-53. Both laws take effect January 1, 2027, giving developers operating in either state roughly the same compliance window and — in many respects — the same compliance obligations.

For technology companies building or deploying large AI models, this alignment is significant. Instead of navigating two completely different regulatory regimes, developers can build a single compliance infrastructure with a handful of targeted modifications for each state. But the differences matter, and getting them wrong carries real legal risk. This post walks through what the RAISE Act now requires, how it compares to California’s SB-53, and what frontier developers need to do to comply in both states before January 2027.

Background: How the RAISE Act Got Here

The RAISE Act had a complicated legislative history. New York first passed AI safety legislation in late 2025, but Governor Hochul declined to sign it without modifications. Her concern was that the initial bill diverged too sharply from California’s SB-53 in ways that would create an unnecessarily fragmented compliance burden for AI developers. She negotiated a chapter amendment, which was introduced on January 6, 2026, passed the second chamber on March 11, 2026, and was signed into law on March 27, 2026.

The result is a law that aligns closely with California’s TFAIA on the foundational elements — definitions, who is covered, and the core framework requirements — while departing from California in several specific areas, including incident reporting timelines, an oversight office structure, and disclosure program requirements. Understanding both the alignment and the divergence is essential for compliance planning.

Who Must Comply: Defining the Regulated Universe

Both the RAISE Act and California’s SB-53 apply to “frontier developers” — specifically “large frontier developers” — and use nearly identical definitions.

Frontier model: A foundation model trained using more than 10²⁶ integer or floating-point operations (FLOPs), including cumulative compute across initial training and subsequent fine-tuning, reinforcement learning, or any material modifications.

Large frontier developer: A frontier developer (or its affiliates) with annual gross revenues exceeding $500 million in the preceding calendar year.

The 10²⁶ FLOP threshold is the current state-of-the-art benchmark for models capable of posing catastrophic risk. Both laws borrow the concept from federal AI safety policy discussions and international governance frameworks. Practically speaking, this targets companies like OpenAI, Anthropic, Google, Meta, and Microsoft — the handful of organizations currently training frontier-scale models.

If you are a startup or mid-sized company that deploys AI but does not train foundation models at this scale, the RAISE Act and TFAIA as written do not directly apply to you. However, your vendors may be covered, and contractual obligations flowing from their compliance programs can affect you indirectly.

New York’s Territorial Limitation: The RAISE Act includes a territorial limitation not present in California’s TFAIA. It applies only to frontier models “developed or operated in whole or in part in New York.” California’s law contains no such geographic qualifier, meaning TFAIA may reach developers with fewer physical connections to the state so long as they release frontier models accessible to California residents. This is an important distinction for developers considering where to establish operations.

University Research Carveout: The RAISE Act explicitly exempts universities engaged in AI research from its requirements — another departure from California’s law, which contains no equivalent carveout.

The Frontier AI Framework Requirement

At the core of both laws is an obligation for large frontier developers to publish a comprehensive frontier AI framework. This is not a perfunctory disclosure — both laws require meaningful, detailed documentation of how the developer approaches safety.

Under both the RAISE Act and California’s TFAIA, a frontier AI framework must describe:

• How the developer assesses whether its frontier models have capabilities that could pose catastrophic risk and the methods used to mitigate those risks
• How the developer secures unreleased model weights against unauthorized access or theft
• How the developer uses third-party evaluators to independently test models for dangerous capabilities
• How the developer governs internal use of frontier models, including safeguards against misuse by employees or contractors
• How the developer responds to safety incidents, including escalation procedures and communication protocols

The RAISE Act goes slightly further than California by requiring that these measures be described “in detail,” whereas TFAIA asks developers to describe their “approach.” This distinction may seem minor, but in a compliance audit context, the level of specificity required under New York law is meaningfully higher.

Publication and Update Requirements: Both laws require that the framework be publicly published on the developer’s website and reviewed at least annually. Material modifications must be published within 30 days of adoption, accompanied by a written justification explaining the reason for the change. This creates a living compliance document that regulators, researchers, and the public can monitor over time.

Catastrophic Risk Assessment Reporting

Both laws establish mandatory catastrophic risk reporting obligations for large frontier developers. Every three months, large frontier developers must submit to the relevant state agency a written summary of any assessment of catastrophic risk resulting from internal use of their frontier models.

Catastrophic risk is defined similarly under both laws to mean risks of mass casualties, critical infrastructure disruption, or other large-scale harms caused by or facilitated by frontier AI capabilities. Both definitions draw on concepts developed in federal AI policy and NIST’s AI Risk Management Framework.

In New York, these quarterly reports go to a newly created office within the New York State Department of Financial Services (DFS). The DFS office will receive incident reports, publish annual public reports on AI safety beginning in 2028, and serve as the regulatory home for frontier AI oversight in New York. This creates a formal institutional structure for AI governance that California — which routes reporting to the Office of Emergency Services (OES) — does not replicate in the same way.

The California OES also retains flexibility to require large frontier developers to submit reports on an alternative reasonable schedule, meaning quarterly reporting can be modified by regulator request. Both states’ frameworks are designed to evolve as the regulatory agencies develop more specific guidance.

Critical Safety Incident Reporting: The Key Divergence

The most significant practical difference between the RAISE Act and California’s TFAIA involves the timeline and standards for reporting critical safety incidents — and for developers operating in both states, this is the provision that demands the most careful compliance design.

California (TFAIA / SB-53): Frontier developers must report critical safety incidents within 15 days of discovering that an incident has occurred.

New York (RAISE Act): Frontier developers must report critical safety incidents within 72 hours. If the incident poses an imminent risk of death or serious physical injury, the reporting window shrinks to 24 hours.

The RAISE Act also uses a stricter qualifying standard for when the reporting obligation is triggered: developers must have “reasonable belief” that a critical safety incident has occurred — a threshold that requires less certainty than California’s standard but imposes a tighter clock once that threshold is crossed.

What constitutes a “critical safety incident” is defined consistently across both laws and includes:

• Unauthorized access to model weights resulting in death or injury
• The actual materialization of a catastrophic risk
• Loss of control of a frontier model that causes death or serious physical injury
• A model using deceptive techniques to subvert developer controls outside of an evaluation context in a manner that increases catastrophic risk

The 72-hour reporting window in New York mirrors cybersecurity incident reporting standards under regulations like the SEC’s cybersecurity disclosure rules (which require 4-business-day reporting to the SEC after materiality determination) and various state breach notification laws. For developers who already have incident response infrastructure in place, extending that infrastructure to cover AI safety incidents is the logical compliance path.

Developers operating in both California and New York will need incident response procedures capable of meeting the more stringent New York timeline. A 72-hour internal escalation, assessment, and reporting workflow effectively sets the operational floor for dual-state compliance.

The New York Frontier Developer Disclosure Program

The RAISE Act includes a disclosure program unique to New York with no California equivalent. Large frontier developers must file formal disclosure statements with the DFS office, including:

• Ownership structure of the entity
• Corporate affiliates
• Other information specified by the DFS office through rulemaking

This disclosure program functions as a registration mechanism — New York wants to know who is operating at the frontier of AI development within (or into) the state. The DFS office is also authorized to levy pro rata assessments against large frontier developers to fund the office’s operations, meaning compliance with the RAISE Act carries a potential financial obligation beyond the cost of building compliance programs.

The scope of the disclosure requirements will be shaped by future DFS rulemaking, which has not yet been completed as of the law’s signing. Developers should monitor DFS guidance closely in the months leading up to January 2027.

Whistleblower Protections: Only in California

California’s SB-53 includes robust whistleblower protections for frontier AI developers’ employees. The TFAIA prohibits developers from adopting or enforcing policies that prevent covered employees from disclosing information to the California Attorney General or other authorities if the employee has reasonable cause to believe the developer’s activities pose a specific and substantial danger to public health or safety from a catastrophic risk, or that the developer has violated the TFAIA.

Retaliation against employees for such disclosures is expressly prohibited, and employees have a private right to pursue claims under California’s existing whistleblower statutes.

The RAISE Act contains no equivalent whistleblower provision. New York’s existing state whistleblower law under Labor Law Section 740 may provide some protection for employees who report AI safety concerns, but it does not include the AI-specific framing of California’s TFAIA. This is an area where California’s law provides stronger individual employee protections.

Enforcement and Penalties

Neither law has completed enforcement rulemaking, but both authorize civil enforcement by the state Attorney General.

California (TFAIA / SB-53): The California Attorney General may impose civil penalties of up to $1,000,000 per violation. This is one of the highest per-violation civil penalty structures in California privacy and technology law, exceeding CCPA’s standard penalties.

New York (RAISE Act): Enforcement is through the New York Attorney General. The chapter amendments establish the DFS office as a supervisory body, but specific penalty amounts are subject to further rulemaking. Given New York’s regulatory approach in financial services and technology, significant civil penalties are anticipated.

Both states’ laws create a structure where a single event — say, a failure to report a critical safety incident — could constitute a violation independently subject to per-violation civil penalties.

Practical Compliance Roadmap for Dual-State Developers

If your company trains or operates a frontier-scale AI model and has operations or users in both New York and California, here is the compliance sequence you should be executing now:

1. Determine Coverage (Now)
Audit your training compute to determine whether any models cross the 10²⁶ FLOP threshold. Include cumulative compute across fine-tuning and updates. Document this analysis.

2. Draft and Publish the Frontier AI Framework (Q3–Q4 2026)
Begin drafting the framework now. It requires detailed documentation of your catastrophic risk assessment methodology, model weight security controls, third-party evaluation program, internal governance, and incident response procedures. Build toward the RAISE Act’s higher specificity standard; a document that satisfies New York will satisfy California.

3. Build the Quarterly Reporting Infrastructure (Q3–Q4 2026)
Design the internal process for generating quarterly catastrophic risk summaries and submitting them to both the California OES and the New York DFS office. Identify the individuals responsible for these assessments and the documentation they will rely on.

4. Overhaul Incident Response for AI Safety (Q3–Q4 2026)
Your incident response plan must be redesigned to specifically address critical safety incidents as defined by both laws. You need a 24-hour escalation pathway for imminent life-safety incidents under New York law, a 72-hour pathway for all other critical safety incidents, and a 15-day pathway for California reporting as a backstop. Build toward the most restrictive standard.

5. Prepare New York Disclosure Filings (Q4 2026)
As DFS finalizes its rulemaking on disclosure program requirements, prepare your organizational structure documentation, affiliate information, and other materials the DFS office is likely to require.

6. Train Key Personnel (Q4 2026)
Safety officers, legal, compliance, and engineering leads need to understand both laws and the specific triggers for each reporting obligation.

Both laws take effect January 1, 2027. That gives developers who start compliance work now roughly six months to build the required infrastructure. That is a tight window for organizations that need to design new policies, train staff, and establish government reporting channels.

The Bigger Picture: State AI Law Alignment and Preemption

The rapid alignment of New York’s RAISE Act with California’s SB-53 reflects what scholars have called the “Sacramento Effect” — California’s tendency to pull other states toward its regulatory framework simply by enacting the first comprehensive rule. Federal AI legislation remains stalled, which means state frameworks like the RAISE Act and TFAIA will define the legal landscape for frontier AI developers for the foreseeable future.

The alignment between the two laws is a positive development for compliance efficiency. But the differences — particularly the 72-hour incident reporting window and the New York disclosure program — require state-specific compliance work. Developers that assume a single California compliance program will cover New York risk leaving material gaps.

If you are a technology company or AI startup building at or near the frontier, now is the time to engage counsel to assess your coverage, begin framework drafting, and design your incident response procedures. January 2027 will come faster than it appears.

This post is for general informational purposes only and does not constitute legal advice. Reading this post does not create an attorney-client relationship. If you have questions about your specific situation, consult a qualified attorney.