On 28 October 2026, Governor Kathy Hochul signed S.9051-B/A.10379-C, adding Article 48 to New York's General Business Law. The Governor's release frames it around chatbots that pretend to be human, build emotional dependency, encourage self-harm or engage in sexually explicit conversations. The statute, which takes effect on 1 January 2027, is much wider than that framing.
The strongest case for the law
The harms that drove the bill are real. Companion-style chatbots are designed to keep users engaged, and a teenager can mistake a system that claims feelings for a friend. Common Sense Media called the targeted features "reckless endangerment of our children." Some of the 12 prohibited features are hard to dispute. These include outputs that promote suicide, self-harm or disordered eating, outputs that encourage secrecy from parents or professionals, and sexually explicit content aimed at minors. A proportionate law could ban exactly these. Existing consumer-protection law was not built for conversational products, and a state legislature moving here is not unreasonable.
What the text actually covers
The problem is the definition. Under §1800(5) of the enrolled bill, a "covered AI companion" is any "generative artificial intelligence system with a natural language interface" that "provides ongoing, adaptive responses to user inputs." It "includes but is not limited to" an AI companion. Nothing in that language requires a product to be marketed as a companion. A general-purpose assistant appears to fit, and TechTimes reads the law as reaching ChatGPT, Claude and Gemini. That is one commentator's reading and no court has tested it. The text does not obviously contradict it.
The list of "unsafe features" in §1800(8) then mixes narrow harms with open-ended ones. Alongside the clear-cut items, it bars outputs that engage in "flattery or sycophancy," outputs "framed as the covered AI companion's personal opinions or emotional appeals," and outputs that "optimize user engagement that supersede the covered AI companion's safety guardrails." Subsection (h) adds any other design feature that simulates companionship, "as identified via regulations promulgated by the attorney general." Sycophancy is a known model failure that developers are working to reduce. It is also not a bright-line category. A model that praises a student's draft is doing something a regulator could call flattery.
The age gate falls on everyone
The prohibition in §1801 is conditional. An operator may not offer the unsafe features to a covered user unless the user is not a covered minor and the operator has used permissible methods to determine that. In practice, for any user whose age the operator cannot establish, the safe default is to switch the features off or to verify age. Three exemptions exist: customer-service tools, efficiency or research and technical assistance tools, and internal business systems. The first two apply only if the system "is unable to respond on topics outside of the specified purpose." A general assistant cannot meet that condition by definition.
The drafters did add privacy protections. Under §1804, operators must offer more than one age-assurance method, including at least one that does not rely on government ID or that preserves anonymity. Information collected for age checks must be deleted immediately after the attempt, apart from legal exceptions. Self-declaration is only accepted when a user declares that they are a minor. These are better than many state age-verification regimes. They still push operators toward collecting age signals from adults who want only an ordinary answer, and they create a new compliance surface for each operator.
Enforcement and reach
The attorney general enforces the law. §1802 allows civil penalties of up to $25,000 per violation, along with injunctive relief, restitution, damages and the destruction of unlawfully obtained data and any algorithm trained on it. It also voids contract terms that shift liability, and lets a court impose joint and several liability on affiliates if corporate structures were used to limit it. Under §1805 the law applies to conduct occurring "in whole or in part" in New York. A per-violation penalty applied to millions of conversations is a large lever, and vague categories give the enforcer wide discretion over what counts as a violation.
There is also an open question about the rulemaking. Section 1803 lets the attorney general write rules, including defining additional features under (h). That means the real scope of the law will be set after enactment, by regulation, not by the legislature.
A narrower path
The pro-innovation objection is not to protecting children. It is to the drafting. A better law would have defined the harm by outcome: sexual content with minors, promotion of self-harm, deceptive claims of being human. It would have limited the age-gate to products designed or marketed for companionship, and kept subjective behavioral terms like sycophancy for guidance and not for penalties. Instead New York has produced a statute where broad definitions, open-ended features and a per-violation fine will push cautious operators to over-restrict or to over-collect age data, with teens and adults alike as the affected users.
Operators have 65 days before the law takes effect. The decisive documents now are the attorney general's rules and the first enforcement actions. If those stay close to the clearly harmful conduct, the law may prove proportionate. If they stretch to flattery, tone and engagement design, New York will have created a general chatbot-regulation regime through a children's-safety bill.