The registration portal went live on November 10. On its face it is a form: name the developer, list the models, name a contact. What makes it more than paperwork is the date attached to it. Large developers of frontier artificial intelligence that operate in New York have until January 1, 2027, to register under the state’s RAISE Act, and once registered they must report a serious safety incident within 72 hours.
New York is not the first government to write rules for artificial intelligence. It is among the first to build a list of the companies doing the most advanced work, and to attach a clock to what they must do when something goes wrong.
The distinction matters because of what registration is not. It is not a license. The state does not approve a model before release, does not review training data, does not sit in on testing. A developer that registers is not certified as safe. It is known, reachable and on the record.
That is a narrower instrument than a permit and a broader one than a press release. A registry produces a list of entities a regulator can call, inspect and, if it comes to that, name. The reporting duty produces something else: an internal process. A company that must report a serious incident within three days has to decide in advance what counts as serious, who makes the call and what the record will look like. Those decisions are made better before an incident than during one.
The first names on the list are the obvious ones. OpenAI and Anthropic are the initial developers covered, according to the state. Both sell frontier models, both publish safety frameworks of their own, and neither is likely to be startled by a New York filing. It is also not the burden a licensing regime would be, because the underlying work — documenting models, tracking incidents, keeping an audit trail — is work both companies already describe as central to how they operate.
Governor Kathy Hochul paired the registry with an institution. Her office said the state will set up an AI safety research body at Binghamton University, and she described registration and audit together as the basis for a national standard rather than a state-level patch.
That claim will be settled by what New York does with the list. A registry with no inspection capacity is a mailing list. An incident rule with no published findings is a filing cabinet. The pressure point is disclosure: aggregated patterns, categories of failure, the names of systems that keep appearing in reports. Regulators who publish nothing teach nobody anything.
There is a second test, quieter. Frontier developers work across borders. A model deployed in New York is the same model deployed in Ohio, London and Singapore, and a state cannot inspect a training run that happened somewhere else or compel a lab in California to change a release decision. What it can do is create a compliance surface that a company’s lawyers will not ignore, because a state attorney general is a real plaintiff and state law is real law.
How the state defines a large frontier developer will determine how long the list gets. Coverage that turns on the scale of a model will catch research labs with few employees and large compute budgets; coverage that turns on revenue will catch the incumbents and miss them. That definition is not in the portal. It is in the statute, and it will be read closely by anyone deciding whether to file.
For smaller developers, the arithmetic is different. A registry entry is cheap. An audit is not. Independent audits need an auditor, a standard to audit against and a scope that can be checked. None of those is available yet in a form a mid-sized company can buy off the shelf, and the cost of assembling them falls hardest on firms with the least revenue to spare.
The timing has a political logic of its own. State governments have moved faster than Congress on technology rules for most of the past decade, and a governor with national ambitions gains more from going first than from going carefully. If Washington later writes federal rules for frontier models, New York’s registry becomes either a template or a redundancy, and which one it becomes depends on whether the state has produced findings worth copying.
Industry response has been muted, which is itself informative. The companies most affected have the most to lose from a hostile regime, and they have chosen to treat registration as administrative rather than existential. That posture holds as long as the registry stays a registry. It would not survive a state investigation that ended in a court filing.
What the 72-hour clock changes inside a company is easier to describe than what it changes in the world. Incident reporting forces a written record, and written records travel. They surface in litigation, in insurance negotiations and in the diligence files of customers deciding whether to build on a model. A state can shape behavior with a reporting requirement even when it never brings a case, because the requirement makes the behavior legible to everyone who reads it later.
The registration deadline is seven weeks away. The harder deadlines are the ones the law does not set: when the first incident report lands, whether the state publishes it, and whether the researchers in Binghamton get access to anything a company would rather keep. Until one of those things happens, New York has built a door and fitted a lock. Nobody has turned the handle.
