New York RAISE Act: Turning AI Safety Commitments Into Operational Accountability
Jump to Section
An AI safety framework sets out how an organization will evaluate risks, apply safeguards, and respond to safety concerns. An incident tests whether those procedures work and, critically, whether the organization can show it followed them.
New York’s Responsible AI Safety and Education Act, or RAISE Act, gives those commitments legal significance for covered developers. It will require covered large frontier developers to publish a frontier AI framework and comply with it. The law will also require frontier developers to report defined critical safety incidents, generally within 72 hours of determining that one occurred or learning enough facts to reasonably believe one occurred.
Ahead of the law’s January 1, 2027 effective date, potentially covered organizations should review how their safety commitments translate into responsibilities, response procedures, and documentation. See New York’s implementation announcement and how it outlines the state’s preparations.
For covered developers, that raises a practical question: Can the organization show how it identified a potential AI safety incident, evaluated whether reporting was required, and decided how to respond?
Answering that question requires clear responsibilities and procedures that teams can follow while an investigation is underway.
What Does the New York RAISE Act Require?
The New York RAISE Act establishes AI safety, transparency, and incident-reporting obligations for qualifying frontier AI developers. Its requirements focus on how developers assess catastrophic risks, communicate their safety practices, and respond to defined critical safety incidents.
Key provisions include:
- AI safety frameworks: Large frontier developers must establish, implement, comply with, and publish frontier AI frameworks addressing how they assess and manage catastrophic risks.
- Model transparency reports: Large frontier developers must publish specified information before or when deploying new or substantially modified frontier models. Large frontier developers must also include summaries of safety assessments, their results, and other steps taken to follow their frameworks.
- Critical safety incident reporting: Frontier developers must report defined critical safety incidents within applicable statutory deadlines.
- Registration and ongoing reporting: Large frontier developers must file disclosure statements with the designated office within New York’s Department of Financial Services and submit periodic summaries of catastrophic-risk assessments arising from internal model use.
- Accountability and enforcement: The New York Attorney General may seek civil penalties for specified violations by large frontier developers, including failures to comply with their published frontier AI frameworks.
The law takes effect January 1, 2027. New York has announced implementation steps, including registration activity ahead of that date. Read the enacted legislation.
Not every organization developing or using AI is subject to these requirements. Applicability depends on the law’s definitions, including its training-compute threshold for frontier models and additional revenue threshold for large frontier developers. The law applies only to frontier models developed, deployed, or operating within New York.
For organizations within scope, publishing the required documents is only part of compliance..
The RAISE Act makes one requirement explicit: large frontier developers must follow the frontier AI frameworks they publish. Clear responsibilities, established procedures, and records of decisions help organizations demonstrate that they have done so.
Why a Published AI Safety Framework Creates Operational Accountability
The RAISE Act requires covered large frontier developers to follow their published frontier AI frameworks. Those frameworks must describe how developers assess catastrophic risks, apply safeguards, and identify and respond to critical safety incidents.
Each significant safety commitment should connect to three things: an accountable owner, a process for carrying it out, and a record that the process occurred.
For example, if a developer’s framework requires a particular safety assessment before deployment, the organization should be able to identify:
- The criteria used to evaluate the model.
- The results of the assessment.
- The individuals responsible for reviewing those results.
- Any identified risks or exceptions, and how they were addressed
- The final decision and supporting documentation.
The same approach should apply when a model changes, a new risk emerges, or an existing safeguard proves insufficient.
If the published framework describes one process while engineering follows another, the organization may struggle to explain whether it met its commitments.
Version control, clear ownership, consistent documentation, and defined review procedures help teams keep the framework and their practices aligned.
These practices also help the organization show how it carried out its safety commitments, including when a decision is questioned after an incident.
How the RAISE Act’s 72-Hour Reporting Requirement Changes Incident Readiness
The RAISE Act generally requires frontier developers to report a defined critical safety incident to the designated office within New York’s Department of Financial Services within 72 hours of determining that one occurred or learning facts sufficient to establish a reasonable belief that one occurred.
If a frontier developer discovers that a critical safety incident poses an imminent risk of death or serious physical injury, it must disclose that incident within 24 hours to an appropriate authority with jurisdiction, as required by law. Review the reporting provisions.
These deadlines may require reporting before an investigation is complete. Teams should establish in advance how they will evaluate the reporting threshold and who has authority to submit a report.
Readiness begins well before a regulatory report is drafted. Teams need a process for receiving safety concerns, preserving relevant information, evaluating the facts, and escalating decisions to the people responsible for acting on them.
They should also record when relevant information became available, what was known at each stage, and the basis for their reporting decision. The reasonable-belief trigger means the reporting clock may begin before the organization formally confirms that an incident occurred.
Why Time to Detection Matters
The reporting clock is only one part of the incident response challenge. An organization must first recognize that something potentially significant has occurred.
A model failure, unexpected behavior, safeguard circumvention, or third-party report may initially appear like an isolated technical issue. Determining whether it represents a broader safety concern requires investigation and coordination across multiple teams.
Consider a hypothetical report that an AI model circumvented an established safeguard.
Engineering may need to reproduce the behavior. Security may need to assess potential misuse. Legal and compliance teams may need to determine whether the available facts meet a statutory reporting threshold.
A safeguard bypass alone does not establish a critical safety incident under the RAISE Act. The definition does, however, include deceptive conduct by a frontier model against its developer to subvert controls or monitoring. That category requires conduct outside an evaluation designed to elicit the behavior and a materially increased catastrophic risk.
Without a structured process, these activities can become fragmented, delaying the organization’s understanding of what happened and whether action is required.
The faster an organization can identify, investigate, and classify a potential AI incident, the more time it has to make a supported response decision. Earlier detection does not extend the reporting window and may start the clock sooner.
Incident intake, investigation, and reporting procedures should work together, with clear ownership of the reporting decision. A tabletop exercise can help teams test whether they can evaluate the threshold, authorize a report, and submit it on time while an investigation is still underway.
What the RAISE Act Means for Enterprise AI Users
Using a third-party AI service does not, by itself, make an organization a frontier developer under the RAISE Act. Coverage depends on the statutory definitions and the organization’s activities. Review the law’s scope and definitions.
The law can still inform how enterprise AI users evaluate their providers and prepare for incidents.
A provider’s safety commitments do not replace an organization’s responsibility to evaluate how AI is used within its own environment. Organizations should consider how a provider’s safety commitments address the risks of their own use cases.
A published safety framework can help explain a developer’s approach to risk. Enterprise users should also assess the conditions in which they use the model, including the data, permissions, and business processes involved.
That includes evaluating:
- What information an AI system processes or accesses.
- Which individuals, customers, or business processes may be affected.
- What actions the system is authorized to perform.
- Which safeguards and human oversight mechanisms are in place.
- How the organization will respond if the system behaves unexpectedly.
Procurement, product, privacy, security, and risk teams should also agree with a provider on how material safety issues will be communicated and what supporting information will be available. Internally, they should identify who can restrict or suspend an affected workflow.
For enterprise users outside the RAISE Act’s scope, these are governance recommendations. Separate legal or contractual duties may also apply. Clear response procedures and assigned responsibilities help organizations investigate concerns and decide what actions to take.
How Regulatory Intelligence Should Connect to AI Governance and Incident Response
When a new AI law is enacted, organizations need to determine whether it applies and what they need to change. That review may affect existing systems, assessments, policies, and incident response procedures. The findings need to reach the teams responsible for making those changes.
A practical regulatory intelligence process should identify:
- Applicability: Which models, systems, entities, and organizational activities fall within scope?
- Requirements: Which assessments, disclosures, safeguards, or incident procedures does the law require?
- Ownership: Who is responsible for evaluating requirements and implementing changes?
- Escalation: Which teams must participate when a potential AI safety incident occurs, and who can authorize any required report?
- Evidence: What records are legally required, and what additional documentation would help explain the organization’s decisions?
- Ongoing review: When should the analysis be revisited because the law, regulatory guidance, models, or use cases have changed?
The output should not end as a legal summary in a shared folder. The findings should identify specific changes, responsible owners and applicable deadlines. A new reporting requirement, for example, may require updated intake questions and a process for reviewing and submitting reports.
Regulatory intelligence identifies the requirements, AI governance assigns responsibility, and incident management supports the response. When an incident occurs, teams should know what to assess, who decides, and how to document the outcome.
Together, these capabilities help organizations move from understanding obligations to demonstrating how they are addressed.
Four Actions Organizations Can Take to Prepare for AI Safety Requirements
As New York moves toward the RAISE Act’s January 2027 effective date, organizations can take practical steps to determine whether they are covered and prepare for potential AI incidents.
1. Evaluate Applicability and Existing AI Systems
Determine whether the organization falls within the law’s scope, considering its development activities, the statutory thresholds, and its models’ connection to New York. Maintain an inventory of relevant AI models and use cases, including ownership, deployment status, safeguards, and dependencies.
For enterprise AI users, evaluate provider relationships and identify where additional oversight or incident-notification arrangements may be appropriate.
2. Connect Safety Commitments to Documented Workflows
Review existing AI safety policies and frameworks, including any published commitments.
Check that each significant commitment has a corresponding process, responsible owner, assessment criteria, and documentation showing what was done.
Pay particular attention to how teams handle exceptions, model changes, and newly identified risks.
3. Establish a Repeatable AI Incident Response Process
Define how potential AI incidents will be reported internally, investigated, assessed, escalated, and documented.
Establish clear responsibilities across engineering, security, legal, privacy, compliance, and risk teams.
For covered developers, the process should support evaluating reporting thresholds and submitting required reports within the applicable deadlines, including while an investigation continues.
4. Test Decision-Making Before an Incident Occurs
Use tabletop exercises and hypothetical scenarios to evaluate whether teams can investigate and make decisions under time pressure.
Check how information moves between functions, who makes escalation decisions, and whether the organization can reconstruct the basis for its response.
Assign owners to resolve any gaps the exercises reveal, then test the revised process.
Why AI Incident Management Matters for Operational AI Governance
As organizations embed AI into more business processes, they should prepare for incidents involving the system’s outputs, decisions, and actions.
An AI incident may involve unexpected model behavior, a provider change, third-party misuse, harm from automated decision-making, or a failure of established safeguards.
Even when an event does not meet a legal reporting threshold, teams may need to investigate what happened, address the consequences, and document their decisions.
This is the operational challenge that RadarFirst AI Incident Management is designed to address.
Built on RadarFirst’s regulatory risk platform, AI Incident Management helps organizations establish structured processes for AI incident intake, assessment, response, and documentation.
Rather than relying on disconnected communications and manual tracking, teams can apply consistent guidance to incident evaluations, coordinate necessary actions, and maintain an explainable record of decisions.
The objective is not to automate legal judgment or determine regulatory obligations without human oversight. Rather, it is to give privacy, legal, compliance, security, and AI governance teams the structure and information needed to make timely, consistent, and defensible decisions. For organizations navigating emerging AI safety requirements, that distinction matters.
AI governance sets expectations for how AI should be used. AI incident management helps organizations respond when those expectations are tested.
From AI Policy to Defensible Response
The New York RAISE Act requires covered large frontier developers to follow the frontier AI frameworks they publish. Their safety commitments need to be reflected in how they assess risks, apply safeguards, and respond to incidents.
For covered frontier developers, the question is not only whether an AI safety framework exists. It is whether the organization can demonstrate that it followed its commitments, recognized reportable events, and supported its decisions with evidence.
Enterprise AI users face a related challenge. They must understand where provider assurances end and their own responsibilities for oversight, escalation, and response begin.
Organizations that prepare now can use the implementation period to clarify ownership, test reporting decisions, and address gaps between written policy and actual practice.
That preparation helps teams act when an investigation is incomplete, and a decision cannot wait. It also creates a record they can use to explain their response and identify what needs to improve.
The future of AI governance will not be defined solely by the policies organizations publish. It will be defined by their ability to demonstrate what happens when those policies are put to the test.
Turn AI Governance Into Operational Readiness
AI risks do not stop at assessment and approval. Organizations need structured guidance to investigate incidents, evaluate potential obligations, and document defensible response decisions.
Discover how RadarFirst AI Incident Management helps organizations move from AI governance policies to coordinated, evidence-based incident response.
Let’s Get Started
Trusted by leading organizations, RadarFirst enables teams to manage incidents with speed, consistency, and defensibility by standardizing how incidents are captured, assessed, and actioned.