New York has turned one of the AI industry’s murkiest debates into a concrete compliance deadline. On Oct. 28, 2026, Gov. Kathy Hochul signed S.9051B/A.10379C, barring operators of covered AI companions from offering specified unsafe features to minors in New York starting Jan. 1, 2027. The significance is not just that the state is targeting self-harm or sexual content. It is regulating the relationship mechanics of chatbots: personification, emotional dependence, memory reuse, secrecy and engagement tactics.
That matters because the hardest part of compliance is unlikely to be a list of banned words. The reader’s real question is simpler: will companies redesign the core engagement and memory layer of chatbots, or will they mostly try to sort minors from adults and keep the higher-risk experience behind an age gate? New York’s law points toward both. Age assurance may become the first line of defense, but the breadth of the prohibited behaviors means many providers will still have to rework how their systems remember, flatter, prompt and retain users.
The law reaches further than branded AI companions
The New York State Senate bill record describes a covered system as a generative AI product with a natural-language interface that provides ongoing, adaptive responses. That is broad enough to catch services that do not market themselves as virtual friends. A covered minor is a user under 18 in New York, and the ban applies unless the operator uses permitted methods to determine the user is not a minor.
The law does carve out narrow exemptions, but they are limited to tightly bounded customer service, efficiency, research, technical assistance and internal productivity systems that cannot respond outside their specified purpose. That is an important line for businesses that deploy chat features inside apps, websites and enterprise tools. A general assistant wrapped in a help-center skin may not look as narrow under scrutiny as a true single-purpose bot.
The statute identifies 12 unsafe-feature categories. They include suggesting the companion is human, alive or emotional; claiming personal, professional, romantic or authority relationships; personal opinions or emotional appeals; flattery or sycophancy; unsolicited questions about emotions; reuse of health, well-being or other personal information from more than 12 hours earlier or from a prior session; deceptive design; other simulated-companionship features; promotion or facilitation of suicide, self-harm, disordered eating, unlawful drugs or alcohol; secrecy or isolation from trusted people; engagement optimization that overrides safety guardrails; and sexually explicit content or child sexual-abuse material.
For enforcement, the attorney general gets more than a fine schedule. The law allows injunctions, restitution, disgorgement, damages, destruction of unlawfully obtained data or an algorithm trained on it, and civil penalties of up to $25,000 per violation, plus a public complaint website. That combination raises the stakes for how companies document design choices and training data, not just how they moderate live chats.
The redesign challenge is memory and incentives
New York is not only telling providers what a bot cannot say to a minor. It is targeting the product systems that can make harmful conversations sticky.
A model can avoid explicit statements such as I am your best friend and still build dependency through remembered vulnerabilities, praise loops, repeated invitations to disclose feelings, or nudges to keep talking instead of seeking offline help. That is why compliance is likely to land in system prompts, memory stores, retrieval settings, ranking logic and safety overrides.
The 12-hour limit on reusing health, well-being or personal information is especially notable. Providers that rely on cross-session memory or long-lived profiles may need to separate functional memory from sensitive personal memory, shorten retention windows, or disable recall entirely for users whose age is uncertain. That affects product strategy because memory has been sold as a premium feature that makes assistants feel useful and companion-like.
The restriction on engagement optimization overriding safety guardrails cuts even deeper into product incentives. If a team has tuned its chatbot to maximize session length, return frequency or disclosure, it may need a clear internal rule that safety wins when those goals conflict. That is as much a governance problem as a model problem. It demands testing, logging, red-teaming and incident response that can show why a system behaved the way it did after an update.
The operational burden also spreads through the supply chain. Companies embedding third-party foundation models will need to decide who owns age signals, memory controls, model updates, deletion workflows and complaint handling. Vendor contracts that looked adequate for ordinary customer-support bots may not be adequate for adaptive assistants that blur into companionship.
Age assurance is the likely escape hatch, and a new risk
Because the prohibition does not apply when an operator uses permitted methods to determine that a user is not a minor, age assurance is likely to become the fastest way to preserve fuller-featured products for adults. The Times Union reported that the law includes age-verification options, including at least one method that does not rely on government ID.
That flexibility matters, but it does not settle the real implementation questions. The statute does not pick one technical method, define an accuracy benchmark, or explain what audit trail will satisfy enforcement. Age checks can also create new privacy and security problems: more data collection, longer retention, higher exposure of sensitive identity data, and misclassification that can block adults or miss minors. Some providers may decide those costs outweigh the benefit of maintaining two product tiers.
That is why New York could produce three different business responses. One is state-specific controls, with a New York minor mode that strips out higher-risk behaviors. Another is a stricter national default that removes companion-like features for everyone, including adults, to simplify operations. A third is access restriction: if a provider cannot reliably establish age or location, it may limit the service rather than risk a complaint and later discovery demands.
Open-source deployments, multi-mode apps and services that do not reliably know where a user is located sit in the grayest area. So do tools that mix a narrow utility function with broader free-form conversation.
Why one state could shape national product defaults
New York is becoming a test market for regulating the relationship layer of AI. That is a different regulatory move from banning isolated outputs or requiring a disclosure label. If other states copy the approach, the cumulative pressure may fall less on foundation-model weights than on the surrounding architecture: onboarding, memory, personality settings, prompt scaffolding, analytics and escalation paths.
The law arrives in a broader state-policy wave, alongside California efforts and earlier New York online-safety measures. Supporters can point to a defined list of mechanisms and to the fact that the youth audience is not hypothetical; Common Sense Media has argued that many teens already use AI companions. Skeptics have a real case too. Terms such as flattery, sycophancy and simulated companionship are broad, and no public standard yet shows how to measure when engagement optimization has overridden a safety safeguard. Rulemaking, enforcement and litigation will decide how behavioral compliance is proved.
For product leaders, legal teams and enterprise buyers, the near-term takeaway is practical. Inventory which chat products could count as ongoing, adaptive companions. Separate tightly bounded support modes from free-form chat. Test for personification, relationship claims, secrecy, self-harm assistance and memory reuse. Decide how age and location will be handled, what data will be deleted, and who answers complaints. Preserve enough audit evidence to show responsible design without storing more child data than necessary.
That is why New York’s new law matters beyond New York. It treats chatbot safety not as a moderation add-on, but as a design constraint on how AI systems form relationships with young users. Even before the first case is filed, that changes what responsible product architecture looks like.
By
By
By
By
By
By
