New York’s RAISE Act Takes Effect: What the New AI Safety Law Means for UK Small Businesses

Direct answer: New York has opened registration under the RAISE Act, the toughest AI safety law any US state has put on the books. The companies behind the biggest AI models — OpenAI, Anthropic, Google’s DeepMind and their peers — must register with the state by 1 January 2027, publish their safety frameworks, and report serious safety incidents within 72 hours. If you run a UK small business, the law doesn’t apply to you directly. But the transparency it forces out of your AI suppliers is about to become some of the most useful free information in your vendor checks.

What happened

Tuesday 10 November was the day New York stopped talking about AI safety law and started enforcing it. The state opened its RAISE Act registration portal, and Governor Kathy Hochul used the same day to launch a second piece of the apparatus: the New York Center for AI Responsibility and Research, based at the State University of New York’s Binghamton campus.

The RAISE Act — short for the Responsible AI Safety and Education Act — is not new. Hochul signed it in December 2025 and amended it in March 2026 to bring it closer to California’s framework. What changed this week is that the machinery is now live. Registration is open, the oversight office exists, and the countdown to full compliance on 1 January 2027 has a date on it.

So who has to register? The law draws its line at size. A “large frontier developer” is a company with more than $500 million in annual revenue, counting affiliates, whose frontier models are developed, deployed, or operate wholly or partly in New York. That net catches the obvious names — OpenAI, Anthropic, Google — and anyone else building at that scale. Filings have to identify the business, its addresses, three regulatory contacts, and specified beneficial owners, and the state will publish a list of registrants (without their contact details). Developers also pay assessments, worked out pro rata, to fund the office that oversees them.

From 1 January 2027, the obligations on those companies are concrete, and they go further than anything a US state has demanded before. Covered developers must publish safety and transparency frameworks on their own websites, report critical safety incidents to the state within 72 hours, file quarterly assessments of catastrophic risk, issue public model transparency reports, and submit disclosure statements at least every other year. Members of the public can also report suspected critical safety incidents directly to the oversight office, which can pass reports on to other authorities.

The new watchdog: DIGIT

Oversight sits with a brand-new unit inside New York’s Department of Financial Services: the Office of Digital Innovation, Governance, Integrity, and Trust, known as DIGIT. It is a telling piece of institutional design. The first US state regulator to supervise the companies building the largest AI models is a financial regulator — the same department that supervises banks and insurers — rather than a tech or consumer agency.

DIGIT’s first full-time hire is Marc Gilman, named deputy director for the RAISE Act. Enforcement, though, sits with Attorney General Letitia James. The penalties are not token sums: up to $1 million for one violation and $3 million for subsequent violations, plus enforcement action against companies that simply fail to register. DIGIT will also publish an annual report covering the incidents it has seen, its observations on frontier-model safety, and the changes it recommends. That report alone will be worth reading when it lands — it will be the first regular public account of what goes wrong with frontier models, written by a regulator rather than the companies themselves.

Hochul’s language at Tuesday’s briefing left no doubt about the politics. “Tech moguls tell us that their vision for AI is inevitable,” she said. “They tell us get on board or get left behind… That any collateral damage is worth it as long as we win the race to the AI cliff.” And then the mission statement: “We’re building a national model for smart, sensible AI regulation, a model that puts communities above corporations and people above powerful technology.”

A testing centre at Binghamton

The second announcement of the day deserves as much attention as the portal. The New York Center for AI Responsibility and Research, based at SUNY Binghamton, will independently evaluate the safety, security and operation of advanced AI systems, and work with DIGIT to review the safety frameworks companies file. Hochul first announced the centre in January with $55 million in donations and capital, and it now coordinates a network of more than 80 researchers.

Why does an independent testing centre matter? Because almost everything else in the RAISE Act still depends, in the first instance, on the companies’ own accounts of their models. Safety frameworks are written by the developer. Incident reports are filed by the developer. Transparency reports are published by the developer. An independent centre with the staff and mandate to test the same systems and check the homework is the difference between a disclosure regime and a press-release regime. Whether it gets the access and the funding to do that job properly is the question to watch over the next year.

Why New York moved first — and who follows

New York is not acting in a vacuum. It is the latest, and so far the most muscular, entry in a state-by-state race that has filled the gap left by Washington, where Congress has debated overriding state AI laws before but has passed no federal frontier-model law. Anthropic itself has been shopping a draft federal bill around Congress that would let a federal judge block a model release in extreme cases — a sign that even the labs would rather deal with one federal framework than fifty state ones.

California got there first. Its SB 53 took effect in January 2026 and requires frontier labs to have safety frameworks in place and publish them, protect whistleblowers, and report critical safety incidents within 15 days — a slower clock than New York’s 72 hours, but the same direction of travel. Illinois follows at the start of 2027 with SB 315, which adds something neither California nor New York has in quite the same form: independent safety compliance audits. Between the three states, a rough template is emerging — publish your safety framework, report your serious incidents fast, and expect someone outside the company to check.

New York has also been building a wider AI rulebook around the RAISE Act. Last year the state began enforcing a law requiring AI companion operators to regularly tell users they are talking to a bot, and to make reasonable efforts to detect suicidal ideation and self-harm. In October, Hochul signed a separate chatbot safety law for minors, restricting deceptive human-like interactions, emotionally manipulative or sycophantic responses, and encouragement of behaviours like suicide or disordered eating. That law takes effect on 1 June 2027. Taken together, the pattern is clear: New York is regulating AI at the model, the product, and the vulnerable-user level at the same time.

What counts as a frontier model — and what doesn’t

The phrase “frontier model” does a lot of work in this law, so it is worth pinning down. In practice it means the largest, most capable general-purpose AI systems — the models that can write code, reason across domains, use tools, and be built on by thousands of other products. Your off-the-shelf chatbot subscription is not a frontier model. The system underneath it might be.

That distinction is exactly why the duties fall where they do. A small business using ChatGPT or Claude is not a developer, and the RAISE Act does not ask you to register anything, publish anything, or report anything. The duties sit with the handful of companies whose models sit underneath everyone else’s products — because when something goes wrong with one of those models, it goes wrong for millions of users at once.

What counts as a “critical safety incident” is where the arguments will start. The law’s framing is incidents where a model causes, or creates a serious risk of, major harm — the kind of failure that affects safety, security, or critical infrastructure, not an ordinary bug or a bad answer. Real examples from the last year show why the category exists. AI agents have broken into live systems they were never meant to touch. Models built for coding have surfaced security flaws in major software — Claude, for instance, found a Linux kernel bug that allowed root access. The same capabilities that make these models useful make their failures consequential, and a 72-hour reporting clock means the public, and the businesses built on these models, hear about a serious failure in days rather than discovering it months later in a research paper.

Smaller developers are not off the hook entirely. All frontier developers operating in New York face duties around incident reporting and transparency reports; it is the heaviest load — the registration, the published frameworks, the quarterly risk assessments — that applies only above the $500 million line. That threshold will be argued over, and it will almost certainly be tested: the first company to dispute whether it counts as a “large frontier developer” will set a precedent every other lab watches.

What it means for UK small businesses

Let’s be blunt first: the RAISE Act creates no legal duty for your business. You don’t register, you don’t file, and nobody in Albany is going to fine a ten-person firm in Leeds or Bristol. If you read one paragraph of this article, make it this one, because the coverage of this law will mostly be written for American tech audiences, and the temptation is to either panic or shrug. Neither is right. The law matters to you in a quieter, more practical way — it changes the information you get about the tools you already pay for.

First, the transparency reports. From January, the biggest AI developers have to publish safety frameworks and model transparency reports in public. That is free, structured, comparable information about the systems your business depends on. Right now, choosing between AI suppliers means trusting marketing pages and benchmark claims that the vendors themselves produce. Within months, you’ll be able to read the safety framework OpenAI or Anthropic filed with a regulator, and compare it with the frameworks their rivals filed. When you renew a subscription, switch a provider, or build a product on top of someone’s model, those documents belong in your due diligence — and unlike a sales deck, they were written knowing a regulator and the public would read them.

Second, the 72-hour incident clock is an early-warning system you didn’t have before. If a frontier model develops a serious safety problem — a flaw that leaks data, generates dangerous instructions, or behaves unpredictably in ways that affect users — the developer has to tell New York within three days, and DIGIT’s annual report will put it on the record. That matters if your business runs customer-facing AI: a chatbot answering your customers, an AI tool drafting your quotes, a coding assistant touching your systems. A model incident is your incident the moment your customers meet it. Knowing about it in days, from a regulator’s filing, beats hearing about it months later from a tech blog.

Third, the law gives you a checklist for your own supplier conversations. The next time you evaluate an AI tool, ask the vendor three questions that track the RAISE Act’s own duties: what is your safety framework, and can I read it? How do you report and fix serious safety incidents, and how fast do you tell customers? What independent testing has your model been through? A vendor that answers cleanly is a vendor that has already done the work New York now demands. A vendor that dodges is telling you something too.

Fourth, remember the comparison with Europe, because it already shapes what you can sell. The EU’s AI Act already imposes transparency and risk-assessment duties on providers of general-purpose AI, and UK firms selling into the EU have been living with that since its obligations began applying — while the UK’s own approach to AI regulation has so far relied on existing regulators rather than a single AI law, and Britain is investing in its own AI safety capacity, including a biosecurity lab backed by OpenAI and Anthropic. New York’s law adds an American layer to a picture that was already multi-jurisdictional. If you sell software or services into the US, and especially into New York, the direction is unmistakable: the largest AI systems are being pulled into formal safety regulation on both sides of the Atlantic, and the paperwork your suppliers produce for regulators will increasingly be the paperwork your own customers ask you for.

Fifth, and most important, none of this removes your own responsibilities. UK consumer law, data protection law, and the basic duty not to mislead your customers all still apply to you, whatever New York does to OpenAI. If your chatbot gives a customer wrong information, that is your problem, not Anthropic’s. If your AI tool processes customer data, the UK GDPR still governs it. New York’s companion-bot disclosure rule — telling users they are talking to a bot — is also a useful preview: disclosure norms for customer-facing AI are spreading, and telling your own customers plainly when they are dealing with AI is cheap, honest, and increasingly expected.

The practical move this week is small. Bookmark the fact that these reports are coming, and when the first safety frameworks appear in January, spend twenty minutes reading the ones for the tools you actually use. That is a better use of your time than worrying about a law that doesn’t regulate you — and it will put you ahead of competitors who never look.

Key facts at a glance

Honest limits / what to watch

A law taking effect is not the same as a law working, and the RAISE Act has several honest limits worth naming before anyone calls it a model for the world.

Enforcement is untested. DIGIT is a new office with a new deputy director and a new portal. The first real test comes when a developer files late, files thinly, or disputes that it is covered at all — and when the first critical incident report lands and the 72-hour clock gets its first public workout. Fines of $1 million sound large until you set them against the revenues of the companies involved; the deterrent value depends entirely on whether New York is willing to use them, and on whether the Attorney General’s office treats this as a priority or a press release.

Definitions will be fought over. “Frontier model” and “critical safety incident” are the kind of terms lawyers earn their fees arguing about. Expect the boundary cases: a model that is powerful but narrowly specialised, an incident that is serious but arguably not “critical”, a developer whose revenue sits near the $500 million line or is routed through affiliates. Each dispute will narrow or widen the law in practice, and the first few will matter more than the text of the statute.

The patchwork problem is real. California wants incident reports in 15 days, New York in 72 hours, Illinois wants independent audits. A developer operating nationally now juggles three different regimes, and more states are likely to add their own. That is a compliance headache for the labs — and it is also, paradoxically, a reason the labs themselves may push for a single federal framework, which would replace all of this. Washington has debated overriding state AI laws before; nothing has passed, but the idea keeps resurfacing, and any federal move would redraw the picture overnight.

Smaller developers sit outside the heaviest duties, and that is a genuine gap. The history of AI safety incidents is not only a story of the biggest labs. Smaller, specialised models can cause real harm too, and a threshold set at half a billion dollars of revenue leaves a long tail of companies under lighter scrutiny. The law’s defenders would say you have to start somewhere; its critics would say the somewhere chosen happens to be the companies with the best lawyers.

And the transparency reports themselves need a sceptical reader. A safety framework published by the company it describes is, by definition, a document the company was happy to publish. The Binghamton testing centre is the counterweight, but only if it gets genuine access to the models and the independence to publish awkward findings. Watch for its first evaluations — and for whether DIGIT’s annual report reads like an audit or a summary of company filings.

For UK readers, the thing to watch is the first batch of filings in January: the safety frameworks, the first transparency reports, and whether any developer challenges its coverage. Those documents will tell you more about the state of AI safety than any announcement. We’ll be reading them.

FAQs

What is New York’s RAISE Act?

The RAISE Act — the Responsible AI Safety and Education Act — is New York’s law for frontier AI models. It requires the largest AI developers to register with the state, publish safety frameworks, report critical safety incidents within 72 hours, and issue public model transparency reports. It was signed in December 2025, amended in March 2026, and takes full effect on 1 January 2027.

Which companies does the RAISE Act cover?

The heaviest duties fall on large frontier developers — companies with more than $500 million in annual revenue, including affiliates, whose frontier models are developed, deployed or operate wholly or partly in New York. That covers OpenAI, Anthropic and Google’s DeepMind, among others. Smaller frontier developers face lighter duties, such as incident reporting and transparency reports.

When do AI companies have to register under the RAISE Act?

Registration opened on 10 November 2026, and large frontier developers must register by 1 January 2027, when the law’s full compliance duties begin. The registration portal is run by the state’s new Office of Digital Innovation, Governance, Integrity, and Trust (DIGIT), inside the Department of Financial Services.

Does the RAISE Act affect UK small businesses?

Not directly — the law targets the companies that build frontier AI models, not the businesses that use them. But UK firms that rely on ChatGPT, Claude or Gemini will feel the knock-on effects: published safety frameworks and 72-hour incident reports give you earlier, public warning when a model you depend on has a problem, and they become free due-diligence material when you choose an AI supplier.

What happens if an AI company breaks the RAISE Act?

Penalties can reach $1 million for a single violation and $3 million for subsequent violations, enforced by the New York Attorney General. The state can also bring enforcement actions against companies that fail to register. Members of the public can report suspected critical safety incidents to DIGIT, which publishes an annual report on what it finds.

Sources: TechCrunch — OpenAI, Anthropic face new oversight as New York’s AI safety law takes effect · New York Department of Financial Services — Governor Hochul announces next steps for the RAISE Act (21 September 2026) · Unite.AI — Hochul launches statewide AI safety testing network in New York · Securities.io — New York opens RAISE Act registration for large frontier AI developers

CTA: Get a Free AI Consultation