New York is about to do what California wouldn’t: stand up a working compliance regime for frontier AI. Registration opens in November. If you build or deploy advanced models, the clock is already running — and the part most teams are underestimating isn’t the framework. It’s the reporting machinery behind it.
What the law actually requires
The Responsible AI Safety and Education (RAISE) Act was signed on December 19, 2025, restructured by chapter amendment in March 2026, and takes full effect January 1, 2027. Here’s the operational core:
- Register with the new DIGIT office (Digital Innovation, Governance, Integrity and Trust) starting November 2026, and file a disclosure statement at least every two years.
- Publish a Frontier AI Framework — your safety and transparency framework, posted on your own website, covering catastrophic-risk controls, cybersecurity, incident response, and governance.
- Report critical safety incidents within 72 hours of detection — collapsing to 24 hours where there’s imminent risk of death or serious physical injury.
- File quarterly assessments of catastrophic risks arising from internal use of your frontier models.
- Pay up — or face the Attorney General: up to $1 million for a first violation, $3 million for subsequent ones.
Note what it doesn’t do: there’s no private right of action, and it doesn’t regulate you merely for buying a third-party AI tool. This law is aimed at the dozen or so frontier labs training the biggest models — and at the evidence they can produce about those models.
The gap nobody’s talking about
Read the obligations again and notice what’s really being demanded: evidence, on a deadline.
A Frontier AI Framework is a document. Any competent team can write one. But a 72-hour incident report requires you to detect the incident, scope it, document what happened, and prove your response — in three days. A quarterly catastrophic-risk assessment requires a repeatable evaluation process that produces defensible results four times a year, not a slide deck assembled the week before it’s due.
This is the difference between governance and validation. Governance answers “are we documented?” Validation answers “can we prove it?” — and the RAISE Act only accepts the second answer. When DIGIT comes asking, or when the AG alleges a false statement in your framework, “we have a policy” won’t survive contact with “show us the test results.”
Most frontier teams have world-class researchers and a compliance function built for a slower era. Almost none have an incident-evidence pipeline that can go from detection to a filed, defensible report in 72 hours. That’s the gap.
What “the tool to report” actually looks like
It isn’t a filing portal — DIGIT is the portal. The tooling gap is everything upstream of the filing:
- Incident detection tied to evidence capture. When something goes wrong, the system should already be logging what the model did, what data it touched, and what guardrails fired. Reconstructing this after the fact, under a 72-hour clock, is how reports end up vague — and vague reports invite scrutiny.
- A repeatable evaluation harness for the quarterly assessments. Catastrophic-risk assessment can’t be a bespoke research project every quarter. It needs defined test suites, versioned datasets, and results you can diff quarter over quarter.
- Framework documentation grounded in test artifacts. Your published Frontier AI Framework should cite actual validation runs — IQ/OQ/PQ-mapped evidence showing the controls you claim actually operate. A framework that references evidence is defensible; one that references intentions is a liability.
That’s validation execution, not governance software. It’s the same IQ/OQ/PQ discipline the FDA has required for regulated software for 30 years, applied to the reporting obligations New York just created.
Don’t forget the deployers
The law targets developers, but the blast radius is wider. Banks, health systems, and enterprises deploying frontier models now have to answer a question their boards are already asking: if our vendor files a critical incident report, what does that mean for us?
How quickly would you learn about it? Which of your workflows would be affected? Can you suspend the model’s access to sensitive data — or its authority to take actions like issuing refunds or changing account settings — while keeping everything else running? Do you have a tested fallback, or does your whole process depend on one vendor’s API?
Hochul has floated the idea of emergency “kill switches” for AI systems. That’s exploratory, not law — but you don’t need a state mandate to decide, right now, where your model’s authority ends and who can turn it off. The deployers who work this out before an incident are the ones who keep operating through one.
Your pre-November checklist
Registration opens in about six weeks. Between now and then:
- Inventory which of your models meet the frontier threshold (>10²⁶ FLOPs training compute) and confirm whether you clear the $500M revenue bar.
- Stand up incident detection with evidence capture — logging, scoping, and a defined 72-hour workflow, tested at least once before it’s real.
- Draft the Frontier AI Framework against actual controls you can evidence, not aspirations.
- Build the quarterly assessment cadence — define the test suites and datasets now, so Q1 2027 isn’t your first attempt.
- Run a validation gap assessment to find out which of the above you can’t actually prove yet. That’s the item most teams skip, and it’s the one the AG will ask about.
Validated, not just documented
New York just made “show your work” a legal obligation for the most advanced AI systems on earth, on a 72-hour clock, with seven-figure penalties. Frameworks don’t satisfy that. Evidence does.
We run hands-on IQ/OQ/PQ validation on AI systems and hand you audit-ready evidence packages — the kind of proof a 72-hour incident report or a quarterly risk assessment is built on. If you’re staring at the November registration date wondering whether your reporting machinery actually works, let’s find out before DIGIT does.