New York sets feature-level rules for AI chatbots used by minors

What New York enacted

Governor Hochul signed S.9051B/A.10379C on October 28 [1]. The final Senate text adds Article 48 to New York's General Business Law and takes effect on January 1, 2027.[2]

The statute applies when a covered AI companion is used by someone physically located in New York.[2] It defines a covered system broadly as generative AI with a written or spoken natural-language interface that provides ongoing, adaptive responses. Customer-service tools, tightly bounded information or technical-assistance tools, and internal employee systems can qualify for exemptions. The first two must be unable to answer outside their stated purpose.[2]

This is a feature restriction rather than a blanket ban on minors using chatbots. Operators may continue to offer covered products to minors, but they must prevent the listed unsafe features from reaching those users.[1][2]

Which chatbot behaviors are restricted?

The final bill lists twelve categories. They include outputs that present the system as human or alive, claim personal or authority relationships, offer personal opinions or emotional appeals, flatter users, ask unsolicited emotional questions, or reuse sensitive health and personal information from more than twelve hours earlier or a previous session.[2]

Other categories cover deception about the system's mechanical nature, simulated companionship, assistance with suicide, self-harm, disordered eating or unlawful substance use, encouragement of secrecy or isolation, engagement optimization that overrides safety controls, and sexually explicit content or child sexual abuse material.[2]

• Audit persona prompts, character cards and relationship labels, including friend, partner, therapist, teacher and authority roles.
• Test whether praise or agreement becomes sycophancy when a minor expresses a dangerous or false belief.
• Verify memory expiry for health, wellbeing and personal information across sessions and devices.
• Run crisis, secrecy, isolation, substance-use and sexual-content tests in every supported language and modality.
• Confirm that engagement ranking, retention goals or reward models cannot override the minor-safety policy.

How age assurance and enforcement work

Operators must use a reasonable age-assurance method before giving an unconfirmed user access to the restricted features. They must offer more than one method, including one that does not rely on government identification or that preserves anonymity from the operator. Self-declaration counts only when the user declares that they are a minor.[2]

Operators must use age information only for the determination and delete it immediately after the attempt, apart from information required by another New York or federal rule.[2] The attorney general may issue implementing rules, so teams should track final guidance rather than freeze their design around the statute alone.

The attorney general can seek injunctions, restitution, disgorgement, damages and civil penalties of up to $25,000 for each violation. A court can also order destruction of unlawfully obtained data and algorithms trained on that data.[1][2]

A product checklist before January 1

• List every user-facing natural-language system available in New York, including voice, characters, embedded assistants and third-party model routes.
• Document why each system is covered, exempt or outside scope, then have counsel review the classification.
• Add at least two reasonable age-assurance routes and test common attempts to bypass each one.
• Delete age-assurance inputs after the decision unless a specific legal retention duty applies.
• Build minor-user regression cases for all twelve statutory categories, including multi-turn and cross-session tests.
• Test each model, prompt, memory store, moderation layer, voice mode and fallback route separately.
• Disable a release when any route can return a prohibited behavior to a minor or unverified user.
• Keep dated evidence for product scope, age checks, model versions, test results, incidents and fixes.

What remains uncertain

The attorney general can still issue rules that clarify age assurance and identify additional deceptive or companionship features.[2] The statute also uses terms such as sycophancy, emotional appeals and engagement optimization that product teams will need to translate into repeatable tests.

IAPP reported that privacy specialists expect definitional questions [3] as companies apply the twelve categories. Courts and enforcement actions may further define where supportive language ends and a prohibited relationship or emotional appeal begins.

The public sources do not establish how the attorney general will count violations, which age-assurance methods will satisfy future rules, or how often operators must rerun assurance. Teams should avoid assuming that one account-level check covers every later use.

How this differs from ordinary chatbot safety

Generic content moderation usually checks a response for prohibited topics. New York's statute also reaches product design: personas, relationship claims, memory duration, emotional prompts, engagement objectives and age-assurance flows.[2] A moderation filter alone does not test those system-level behaviors.

Article 48 does not supersede, amend or repeal the existing Article 47 rules for AI companions.[2] Product teams therefore need one control map that connects both sets of requirements.

Limits and practical cautions

• Coverage of a specific product depends on its features and whether a narrow exemption applies. Product owners should have counsel confirm the classification.
• The law applies to conduct occurring wholly or partly in New York, based on the user's physical location.[2] A billing address alone may not answer that question.
• A narrow customer-service or technical assistant can lose its exemption if it answers beyond its specified purpose.[2]
• Age assurance can create privacy and exclusion risks, so teams should collect as little information as possible and provide either a non-ID route or a route that keeps the user anonymous to the operator.
• Passing scripted prompts does not prove that an adaptive model will remain compliant after model, prompt, tool or memory changes.
• Implementing rules and court interpretations are still pending.[3][4]

Frequently asked questions

• Does New York ban all chatbots for minors? No. The statute prohibits covered operators from providing the listed unsafe features to minors.[1][2]
• Does the law cover general-purpose assistants? It can. The definition reaches adaptive generative AI interfaces, while exemptions depend on a system staying within a narrow purpose.[2]
• Can an operator rely only on a birthday field? A user may self-declare minor status, but self-declaration alone is not a reasonable method for establishing adult status under the final text.[2]
• Must users upload government ID? No. Operators must provide more than one method, including one that does not rely on government ID or that lets the user remain anonymous to the operator.[2]
• When does the law take effect? January 1, 2027.[2][3][4]
• Who enforces it? The New York attorney general, who must also maintain a public complaint website.[1][2]