New York AI Safety Law: What Developers Must Prepare13 min read2,662 words

New York AI Safety Law: What Developers Must Prepare

New York’s RAISE Act takes effect in 2027. Learn who must register, what safety disclosures are required, and how AI developers can prepare for oversight.

New York AI safety lawRAISE Actfrontier AI regulationAI safety reportingAI transparency requirements
New York AI Safety Law: What Developers Must Prepare

New York's RAISE Act Makes AI Safety an Operating Requirement

For frontier AI developers, safety promises can't stay stuck on a company website. New York is moving to enforce its RAISE Act, which requires certain large AI developers to register with the state, publish safety frameworks and model transparency reports, and notify regulators of critical safety incidents within 72 hours.

The registration portal is open. Covered developers must register by January 1, 2027. That means companies have to figure out quickly whether they're in scope, who owns each required disclosure, and how incident reporting can actually happen on time. The law also creates a state office to oversee implementation, with penalties of up to $1 million for one violation and $3 million for later violations.

The real question isn't just what the statute says - it's how a company turns those obligations into repeatable processes. This article walks through what the RAISE Act requires, who should pay attention, how it fits into New York's broader regulatory approach, and what AI businesses can do before the deadline.

What does New York's RAISE Act require?

The RAISE Act is New York's safety and transparency law for large frontier AI developers. In practice, it requires companies in scope to register, make parts of their safety work public, report critical incidents to the state, and publish model transparency information.

Even if a developer already runs internal evaluations or has escalation procedures, the law still pushes some of that work outside the organization. Certain safety information must be published, and critical incidents must be reported to the state on a defined schedule.

The three central requirements

The law lays out three expectations for covered developers:

  • Register with New York. A registration portal has opened, and large frontier AI developers must register by January 1, 2027.
  • Publish safety frameworks. Developers must post their safety frameworks on their websites, moving safety documentation from internal governance into public disclosure.
  • Report critical safety incidents within 72 hours and publish model transparency reports. The 72-hour deadline makes incident handling time-sensitive, and the transparency reports add a separate public channel for model-related information.

The source information does not spell out every report format or the exact scope of "critical safety incident." Companies shouldn't try to fill those gaps purely with assumptions. Instead, they should watch for state implementation guidance and set up an internal process that can assess incidents quickly enough to meet the reporting window when something potentially falls within scope.

Why the reporting window changes the work

A 72-hour deadline isn't only a legal issue. If a serious incident is first identified by a product, trust and safety, security, or infrastructure team, that team has to know how to escalate it to the people who make regulatory decisions. If reporting can't begin until after a full internal investigation, the company may miss the window.

That said, not every operational problem automatically qualifies as a reportable critical incident. The key is having a clear path from detection to assessment: named decision-makers, defined steps, and a documented record of what the company concluded. The cleaner the handoff, the less time the organization spends arguing about ownership while a deadline is running.

Who must register, and when?

The RAISE Act applies to large frontier AI developers. The source describes covered companies as those with more than $500 million in annual revenue. That threshold is a useful starting point, but it's not a substitute for checking how the law applies to a particular company or product.

The registration deadline is January 1, 2027. With the portal already open, companies that might meet the threshold have a reason to review their situation before the clock forces a scramble. Registration is only one part of the obligation set - still, it's also a practical trigger for checking whether the company can support the required disclosures and incident reporting.

Revenue is a screening question, not the whole decision

A company can use the revenue threshold to decide whether it needs more digging. It should also identify the specific legal entity that develops or operates the relevant AI systems, who controls the safety framework, and where incident information is actually collected.

This matters more in real-world org charts than people like to admit - subsidiaries, product lines, and shared model-development teams can all complicate who owns what.

A useful internal screening record should cover:

  • The entities and business units that develop or deploy frontier AI systems.
  • The relevant annual revenue figures and how the company calculated them.
  • The teams responsible for safety frameworks and model transparency reports.
  • The systems used to identify and escalate safety incidents.
  • The person responsible for registration and communication with state authorities.

This isn't legal advice. It's a way to avoid a common failure mode: treating a corporate-level question as if it has one obvious owner, when the work is spread across product, research, compliance, security, and operations.

The deadline should shape the sequence of work

Registration is due at the start of 2027, but companies shouldn't wait to start compliance planning until that date. Public safety frameworks and transparency reports require decisions about what the organization can actually disclose. Incident reporting requires a working process, not just a policy statement.

A sensible sequence is to confirm coverage first, assign accountable owners second, and then review the evidence and workflows behind the disclosures. That order reduces the temptation to polish public-facing language before teams know whether it matches how things really work. It also gives leadership time to resolve cross-department gaps, like inconsistent incident definitions or unclear escalation authority.

How will New York enforce the law?

New York created the Office of Digital Innovation, Governance, Integrity, and Trust - DIGIT - to implement the RAISE Act and oversee large frontier AI labs. The state can impose penalties of up to $1 million for one violation and up to $3 million for later violations.

New York can also pursue enforcement actions against companies that don't register through the new portal. A spokesperson declined to provide specifics about how enforcement would work. Businesses should separate what's already described from details that haven't been clarified publicly yet. Uncertainty isn't a reason to ignore the requirement - it's a reason to track guidance and adjust internal processes as implementation becomes clearer.

Oversight includes an independent research center

Along with the registration portal, Governor Kathy Hochul launched the New York Center for AI Responsibility and Research at SUNY Binghamton. The center is meant to independently evaluate the safety, security, and operation of advanced AI systems and work with DIGIT to review companies' safety frameworks.

Hochul announced the center in January, along with $55 million in donations and capital. For companies, the takeaway is straightforward: the safety framework you publish needs to be coherent and grounded in real governance. Outsiders assessing it will look for evidence that the framework matches actual practice, not just good intentions.

Hochul framed the approach as putting "communities above corporations and people above powerful technology." That's political language, but the operational message is practical: New York wants safety practices to be more inspectable than optional.

What the penalties mean in practice

A maximum penalty isn't the same thing as an automatic penalty every time a company slips. The source doesn't explain how violations will be evaluated, what mitigation might be considered, or how enforcement decisions will be made. Companies shouldn't read the headline numbers as a precise forecast of how regulators will act.

Still, the penalty structure gives a clear reason to treat registration, reporting, and publication as managed business processes. Miss a deadline or publish disclosures that can't be supported by internal records, and the problem can quickly become a governance issue - with financial and reputational consequences. The right response isn't grand promises. It's clear ownership and solid documentation for decisions and disclosures.

What should AI companies do before January 2027?

The most useful preparation is linking legal requirements to the teams and systems that already handle model development, deployment, and risk. A policy document doesn't satisfy a 72-hour reporting obligation if nobody knows who can classify an incident or approve a notification. Likewise, a published safety framework creates risk if it claims controls the teams don't actually use.

Companies should create a small compliance workstream with an accountable executive owner, a legal lead, and operational representatives from AI safety, security, product, and communications. The goal isn't a new bureaucracy - it's alignment. The workstream should confirm that registration, disclosures, and incident decisions follow a real path through the organization.

A practical readiness checklist

Before the deadline, companies that may be covered should:

  1. Determine whether the law may apply. Review the revenue threshold and the company's frontier AI activities, and get legal analysis if coverage is unclear.
  2. Complete registration planning. Identify the responsible entity, gather required information, and assign someone to manage the portal submission and any follow-up.
  3. Inventory safety documentation. Compare the published (or planned) safety framework with actual review, testing, and escalation practices. Flag claims that can't be substantiated.
  4. Define incident triage. Set out who receives alerts, who decides whether something might be critical, who makes the reporting decision, and how the company records that decision.
  5. Test the 72-hour workflow. Run a tabletop exercise with the relevant teams to find delays in handoffs and approvals before a real event puts pressure on the process.
  6. Plan model transparency reports. Assign an owner, identify which teams hold the relevant information, and set a review process for accuracy and consistency.
  7. Track state guidance and change. Assign someone to monitor implementation updates and update internal procedures when requirements become clearer.

Keep public commitments connected to evidence

A safety framework works best when it reflects how decisions are actually made. If the document says incidents are escalated through a defined review process, teams should be able to show the process exists and know how to use it. The same principle applies to model transparency reports: the people who own the underlying information should review the public statements, not just communications staff drafting in isolation.

That's where a lightweight evidence trail helps. Keep records of framework reviews, incident assessments, reporting decisions, and approvals. These records don't guarantee compliance, but they make it easier to explain what the company did, why it did it, and whether it followed the process consistently.

For smaller or fast-growing teams, the trade-off is real. Formalizing every decision can slow product work. But for covered systems and high-impact incidents, a narrowly scoped process is often less disruptive than reconstructing what happened after the fact. The point isn't paperwork for its own sake. It's a reliable response when both the stakes and the clock are high.

How does the RAISE Act compare with other state laws?

New York isn't the only state moving toward more direct oversight of AI safety and transparency. California and Illinois have adopted laws with overlapping goals, but they differ in timing and specific provisions. If you operate across state lines, you'll need to understand what's genuinely shared - and what isn't.

California and Illinois take related approaches

California's SB 53, which took effect in January, requires frontier AI labs to establish and publish safety frameworks, provide whistleblower protections, and report critical safety incidents within 15 days. Since that timeline differs from New York's 72 hours, companies shouldn't assume one state's schedule automatically satisfies the other.

Illinois' SB 315 is scheduled to take effect at the start of 2027. It includes related requirements and also calls for independent safety compliance audits. That audit provision is a distinct mechanism and shouldn't be confused with New York's plan to use an independent research center to evaluate advanced AI systems and work with DIGIT.

The state-by-state picture is still evolving. Some incident intake and safety documentation infrastructure can be shared, but each jurisdiction still needs the right rules applied to the right deadlines.

Build a common system, then configure for each state

A practical multi-state approach is to centralize the core work without pretending the legal requirements are identical. Keep one incident intake process, one controlled repository for safety documentation, and clear ownership for public disclosures. Then configure the jurisdiction-specific rules for registration, reporting windows, audits, and other obligations.

This avoids two unhelpful extremes. Duplicating every process state by state creates overhead and inconsistent records. Treating all state laws as one generic "AI compliance" standard can make companies miss important differences - like New York's 72-hour window compared with California's 15 days.

New York has also passed separate laws that focus on consumer-facing AI. A law that began enforcement last year requires AI companion operators to regularly tell users they are speaking with a bot and make reasonable efforts to detect suicidal ideation and self-harm. Another law, signed in October, restricts certain chatbot behaviors involving minors, including deceptive human-like interactions, emotionally manipulative or sycophantic responses, and encouraging suicide or disordered eating. That law takes effect June 1, 2027.

These measures target different products and risks than the RAISE Act. Together, they reinforce why companies should map obligations by system and use case, rather than assuming a single company-wide safety document covers everything.

Conclusion

New York's RAISE Act makes registration, public safety documentation, model transparency, and rapid incident reporting part of the operating environment for large frontier AI developers. The January 1, 2027 deadline is close enough that companies potentially in scope should assign owners and start reviewing their current processes now.

The most reliable approach is practical: confirm whether the company is covered, make sure public claims match real controls, and test whether an incident can move from detection to a reporting decision within the required window. Build shared systems where that reduces duplication - but keep the state-specific rules that change the outcome. That's how compliance becomes part of responsible operations, not a last-minute paperwork exercise.

Frequently Asked Questions

The RAISE Act takes effect at the start of 2027. The registration portal has opened, and large frontier AI developers described as having more than $500 million in annual revenue must register by January 1, 2027.
Covered developers must publish safety frameworks on their websites and issue public model transparency reports. The law also requires registration and reporting of critical safety incidents within 72 hours. Companies should monitor state implementation guidance for details about required formats and definitions.
Penalties can reach $1 million for one violation and $3 million for subsequent violations. New York can also bring enforcement actions against companies that fail to register, though the state has not provided further specifics about that process in the source information.
New York requires critical safety incidents to be reported within 72 hours. California’s SB 53 sets a 15-day reporting window. Companies operating in both states should configure their processes to meet the applicable deadline rather than assume one timeline works everywhere.

Related articles