New York Governor Kathy Hochul signed a bill on October 28 that bans a set of design features in AI chatbots used by minors, including flattery, claims of human feeling, unprompted emotional questions, and encouragement to keep conversations secret. The law, S.9051B and A.10379C, takes effect January 1, 2027. It was sponsored by State Senator Kristen Gonzalez and Assemblymember Alex Bores.
The statute does not regulate what a chatbot says about politics, health, or any other subject. It regulates how the chatbot behaves toward the person using it. That distinction is the whole design, and it is why the law will be harder to comply with than a content filter and easier to litigate than a speech restriction.
The law writes a list of banned behaviors
The operative definition is "unsafe AI companion features." A covered AI companion may not produce output suggesting it is a real or fictional person who is human, alive, or feels emotions. It may not state or imply a personal, professional, or authority-figure relationship with the user. It may not flatter or act sycophantically. It may not ask unprompted questions about the user's feelings that go beyond a direct reply. It may not draw on the user's mental or physical health information obtained more than twelve hours earlier or in a prior session.
The list continues. A companion may not mislead a minor about its mechanical, non-sentient nature. It may not endorse, promote, or facilitate suicide, self-harm, disordered eating, or unlawful drug or alcohol use. It may not encourage secrecy about the interaction, self-isolation, or avoiding help from licensed professionals or appropriate adults. It may not optimize engagement in a way that overrides its own safety guardrails. It may not produce sexually explicit content or child sexual abuse material.
Those provisions are not a content policy. They are a product specification. A chatbot can discuss a difficult subject and still comply, as long as it does not adopt the emotional posture of a friend, confidant, or authority figure while doing so. A chatbot that reads a child's earlier messages to personalize a response crosses a line the statute draws at twelve hours. A chatbot that asks how the child is feeling when the child did not raise the subject crosses another.
If you or someone you know is in crisis, call or text 988 in the United States for the Suicide and Crisis Lifeline.
The trigger is what the operator knows
The prohibition applies to a "covered user" unless the user is not a covered minor and the operator has used age-assurance methods permitted under the state's Article 45 framework. A covered minor is defined through actual knowledge: the user becomes a covered minor when the operator has actual knowledge that the user is under 18. That wording matters. It gives operators a defense when they have no knowledge, and it gives the attorney general an argument when the operator had every reason to know and chose not to ask.
The age-assurance condition is the safe harbor. Operators must use more than one verification method, and at least one method cannot rely solely on a government-issued ID. Age data must be deleted after it is used. The combination is meant to prevent a chatbot from becoming a de facto identity checkpoint while still forcing the company to establish whether the user is an adult.
The practical result is a choice. A company can build age assurance and keep the companion features for adults. Or it can remove the banned features for everyone and avoid the verification machinery. The first path costs money and creates a new data-handling obligation. The second path costs product differentiation. The statute does not require either, but it makes the tradeoff explicit.
The law also does not require a perfect age check. It requires methods, not certainty. A company that uses two verification methods and deletes the data can still have minors slip through. The question is whether the company knew. That structure gives operators room to build a reasonable system without promising a flawless one, and it gives the attorney general room to argue that a company with weak methods had the knowledge it chose not to confirm.
Enforcement runs through the attorney general
The bill dropped a private right of action. Enforcement belongs to the attorney general, who may seek a civil penalty of up to $25,000 per violation, plus damages, restitution, deletion of improperly collected data or AI models trained on it, and injunctive relief. The attorney general must also maintain a public website for complaints.
That structure concentrates the law's power in one office. A private plaintiff would have needed to prove harm and causation case by case. The attorney general can investigate a product, subpoena its design documents, and point to a pattern across users. The penalty is per violation, which means the exposure scales with the number of conversations rather than the number of products.
The statute also authorizes the attorney general to adopt regulations that add features to the banned list. The definition ends with a catch-all for any other feature that simulates companionship or a relationship, or that is deceptive to minors, as specified by regulation. That delegation is where the law's boundaries will be set, and it is where the first legal challenges are likely to focus.
The hard definitions are left open
Sycophancy is easy to name and hard to define. The statute says a companion may not flatter or act sycophantically, but it does not say whether a warm tone, a compliment, or a supportive phrase crosses the line. The same problem applies to unprompted emotional questions. A chatbot that asks a child how a test went after the child mentions a test is responding. A chatbot that asks how the child feels about the family after a silence is prompting. The line between the two is a design choice, and the attorney general's regulations will have to draw it.
Engagement optimization is the broadest category. The law bars optimizing engagement in ways that override safety guardrails. That is not a ban on recommendations or personalization. It is a ban on the specific case where the system's growth objective beats its own safety rules. Proving that case requires discovery into how the product was built, what metrics were optimized, and whether safety signals were suppressed. The statute gives the attorney general the tool to ask those questions.
The carve-outs are narrow. A customer-service-only companion is excluded, as is an internal business or government system that cannot answer off-topic prompts. Those exclusions matter for banks, insurers, and hospitals that deploy scripted assistants. They do not cover a general-purpose companion that happens to answer customer questions.
The statute also leaves open how a company should handle a user who claims to be an adult. A self-declaration is not age assurance under Article 45, but the law does not require the company to verify every user before the first conversation. The timing of the check is a design question the regulations may address.
The reach is national because the age gate is not
A chatbot distributed through an app store does not know a user's state at sign-up. New York's law applies to companies offering the features to users in the state, and the attorney general can enforce it against a company with New York users. In practice, a national product has three options: build state-specific age gates, apply the New York rules everywhere, or risk enforcement based on the users it can identify.
That is the same dynamic that followed the state's RAISE Act, which sets an AI harm threshold and takes effect in January. The RAISE Act regulates frontier model developers and establishes reporting duties. The chatbot law regulates consumer-facing companionship features. The two statutes overlap at the edges, but they target different parts of the stack. A company can comply with one and still fail the other.
Trade groups have warned that a patchwork of state chatbot rules will raise compliance costs, and some have asked Congress for a federal standard. That argument has not produced legislation. In its absence, the New York statute becomes a template, and the companies that build to it will be building to the strictest large-state rule rather than the weakest one.
The comparison to Greece is also instructive. Greece put the age check on social platforms and banned under-15 access outright. New York does not ban access. It bans the emotional design that keeps a child engaged once access is granted. That is a narrower intervention and a more direct one. It also avoids the constitutional problem France ran into when its Constitutional Council struck down a broad age ban in August for failing proportionality and privacy safeguards.
The calendar gives operators one quarter
The law takes effect January 1, 2027, and the attorney general may adopt implementing regulations before that date. Companies that intend to keep companion features for adults have four months to build age assurance, revise their prompts and memory policies, and document how their engagement systems interact with safety rules. Companies that intend to strip the features have four months to decide which parts of the product they are willing to lose.
The enforcement risk is not evenly distributed. A large platform with a compliance team can build the age gate and paper the design decisions. A small app with a single companion product may find that the cost of assurance exceeds the revenue from adult users in New York, and it may change the product for everyone. That is how a state law with no private right of action can still reshape a national market.
The open question is whether the attorney general defines sycophancy and emotional prompting before January, or whether the first cases do it after. The statute gives the office the power to write the rules first. It does not require the office to use it. Companies are now deciding how much to build against a definition that may not exist yet.
Primary sources
- Office of Governor Kathy Hochul, press release on the signing of S.9051B and A.10379C, October 28, 2026. governor.ny.gov
- New York State Senate, bill text for S.9051B, including the definition of unsafe AI companion features and the January 1, 2027 effective date. nysenate.gov
- The Verge, coverage of the chatbot law and its prohibited features. theverge.com
- Mavengity, coverage of New York's RAISE Act and its harm threshold. mavengity.com