The default is the policy
Coverage of the signing has framed this as a companion-app law, and the political presentation pointed at chatbots posing as friends, boyfriends and girlfriends. The operative text is wider than the presentation.
Section 1801 does not ask what an operator knows about a user's age. It makes the 12 features unlawful for a covered user unless two things are both true: the user is not a minor, and the operator has used a permissible method to determine that. An operator that has run no check has met neither condition in a usable form. The practical effect is that an operator cannot offer the listed features to a user whose adulthood has not been established through a permitted method. An unverified adult may therefore face restrictions intended to protect minors until the operator completes an age determination that satisfies the law.
The operative text defines a covered AI companion by its functionality, including a natural-language interface and ongoing, adaptive responses. That definition may reach beyond products marketed explicitly as companions, but its application to general-purpose assistants depends on the enacted wording and how the law is interpreted.
That design is deliberate, and it is where the cost lands. Age assurance adds friction to sign-up and to anonymous use, and the law restricts how operators can soften it. They must offer more than one method, at least one of which avoids government ID or preserves anonymity toward the operator. They may keep the resulting age data for no purpose beyond the determination, and must delete it immediately afterwards. A check designed to be reused for other purposes sits uneasily with those conditions.
Twelve items, several of them ordinary product features
Some items on the list describe conduct no serious operator defends: promoting suicide or self-harm, encouraging a child to keep secrets or avoid trusted adults, sexually explicit material. Those will not be the contested cases.
The more difficult cases involve categories such as flattery or sycophancy, outputs framed as the system's personal opinions, emotional appeals, and certain unprompted questions about a user's feelings. Depending on how the statutory language is interpreted and implemented, these provisions could affect ordinary conversational behaviors in general-purpose assistants. The final boundary will depend on the enacted definitions and any implementing rules or enforcement decisions.
One provision restricts the use of certain health, well-being and personal information acquired more than 12 hours earlier or in a previous session for users whose adulthood has not been established. The provision raises questions for systems that use cross-session memory, although its exact application depends on the statutory wording and any applicable exceptions.
Justine Gluck of the Future of Privacy Forum, speaking to IAPP, put the drafting problem plainly: the categories are “broadly worded and somewhat undefined,” and “it is unclear where ordinary conversational responsiveness ends and prohibited behavior begins.” The Computer and Communications Industry Association made a version of the same argument in asking Hochul to veto the bill, and Tech: NYC criticised its “vague language” after signing. OpenAI and Anthropic did not respond to AFP's request for comment. Until the attorney general issues implementing rules, operators are pricing that ambiguity themselves, feature by feature.
A second law on top of the first one
New York is not starting from zero. Article 47, in force since November 5, 2025, already requires companion systems to disclose that they are programs at the start of an interaction and at least every three hours, and to run self-harm detection and crisis-referral protocols. Its penalties run to $15,000 a day.
Article 48 does something different. Disclosure does not cure a design prohibition. A system can tell a teenager every three hours that it cannot feel emotion and still breach the new law by flattering that teenager, or by drawing on what they disclosed about their health in a previous session. Operators now face both regimes at once: one governing what the system says about itself, the other governing what it may do at all for a user whose age is unresolved.
California has pursued a different mix of protections for minors using AI companions, including requirements concerning safety, disclosures and parental involvement. The two states' frameworks differ in both their regulatory mechanisms and the behaviors they address. Which approach produces fewer bad conversations is an open question. Which one is easier for a small operator to comply with is not: a list with undefined terms and a $25,000 per-violation penalty, enforced by a single office that can also add to the list by regulation, concentrates discretion in the attorney general.
What happens in the next 65 days
The law takes effect on January 1, 2027. The attorney general is responsible for implementing and enforcing its requirements, including the regulatory and complaint mechanisms specified in the statute. Those rules can also extend the banned list under the open-ended item, so the compliance target can move before the deadline arrives.
Operators have the same window to decide three things: which of their features fall inside the 12 items as they read them, which age-assurance method they can offer that meets the deletion and anonymity conditions, and whether their customer-service or research deployments genuinely cannot respond outside their stated purpose. That last condition is stricter than it reads. A support bot built on a general model that can be steered into open conversation does not obviously qualify.
The law may face a First Amendment challenge. Some of its restrictions concern the content or framing of chatbot outputs, raising questions about how they apply to expressive interactions and whether they can survive constitutional review. No challenge had been identified in the reporting reviewed for this article by the time of writing. Whether litigation occurs, and whether a court grants preliminary relief, remain open questions.
Sources
- [1] Bill text of S9051-B/A10379-C (Article 48, Sections 1800-1805), New York State Senate legislation page, verified first-hand for this article: the covered-companion definition in Section 1800(5), the 12 unsafe features in Section 1800(8)(a)-(l), the two-condition prohibition in Section 1801, the exemptions in Section 1801(2), enforcement in Section 1802, and the age-assurance conditions in Section 1804Read source
- [2] New York State Assembly legislative record for S09051B (sponsor Gonzalez; same-as A10379-C; Senate passage June 4, 2026; Assembly passage June 5, 2026; delivered to the governor and signed as Chapter 368 on October 28, 2026)Read source
- [3] IAPP, "New York's governor signs youth AI chatbot safety bill," October 28, 2026 (signing, multiple age-assurance methods, non-ID method, deletion duty, complaint portal, $25,000 penalties, the Gluck/Future of Privacy Forum comparison with California SB 3001, and Hochul's signing remarks)Read source
- [4] AFP via The Business Standard, "New York passes law restricting children's use of AI chatbots," October 29, 2026 (Hochul's "no longer" remarks, Tech: NYC's "vague language" criticism, and OpenAI and Anthropic not responding to AFP)Read source
- [5] MLex, "New York bars chatbots from playing friend or romantic partner to kids," October 28, 2026 (summary of the governor's official statement: features, verification, deletion, penalties, damages and destruction of unlawfully collected data or models trained on it)Read source
- [6] PPC Land, "Hochul signs law barring human-posing AI chatbots for under-18s," October 28, 2026 (bill coverage and reproduction consulted alongside the official text)Read source