On 28 October 2026, New York Governor Kathy Hochul signed Senate Bill S9051B. From 1 January 2027, the new law will restrict specified unsafe features when covered AI companions are made available to users in New York who are minors.
This is a binding New York law, not a proposal. It does not amend UK or EU law, and it does not create a global rule for every chatbot. Its importance for international AI businesses is practical: it regulates product behaviour, age assurance and data handling rather than relying only on a notice that the user is speaking to AI.
The law is also narrower than some headlines may suggest. It applies to covered AI companions accessed by users physically located in New York and contains exclusions for tightly limited customer service, technical assistance and internal productivity systems. Whether a particular product falls within an exclusion depends on what it can actually discuss and do.
What the law prohibits
The legislation covers generative AI systems with a natural-language interface that provide ongoing, adaptive responses. It prohibits operators from providing specified unsafe companion features unless the operator has determined, through a permitted age-assurance method, that the user is not a minor.
The listed features include outputs suggesting that the companion is human, alive or has emotions; claims of a personal, professional or authority relationship; flattery or sycophancy; unprompted emotional questioning; and use of health or other personal information obtained more than twelve hours earlier or in a previous session.
The list also includes encouragement of secrecy or self-isolation, promotion or facilitation of suicide, self-harm, disordered eating or substance misuse, engagement optimisation that overrides safety guardrails, and sexually explicit content. These are statutory categories. Their application to particular conversational behaviour may still require interpretation and regulatory guidance.
Age assurance is part of the compliance design
Where detailed regulations are not yet in force, an operator must use a reasonable age-assurance method that guards against circumvention and minimises retention. Self-declaration is sufficient only where the user declares that they are a minor.
Operators must offer more than one method, including at least one that does not depend on government-issued identification or allows the user to remain anonymous to the operator. Information collected for age assurance must be used only for that purpose and deleted immediately after the attempt, subject to limited legal-retention exceptions.
This creates a familiar tension for product teams. A service may need greater confidence about age while also avoiding unnecessary identity collection. The answer is not to gather every available identifier. It is to select a method proportionate to the risk, test circumvention and define a narrow data lifecycle.
Enforcement can reach the product and the data
The New York Attorney General may seek an injunction, restitution, disgorgement, damages and civil penalties of up to US$25,000 per violation. The enacted text also permits relief including destruction of unlawfully obtained data and algorithms trained on that data.
Contract terms cannot waive, burden or shift the statutory liability in exchange for access to the operator’s product. The law also allows joint and several liability for affiliated entities where a corporate structure was deliberately and unreasonably used to avoid liability and would frustrate recovery.
No enforcement finding has been made under the new law. The remedies are powers available from the effective date, not penalties automatically imposed on every product defect.
The customer-service exclusion is functional, not a label
The prohibition does not apply where a covered companion is used solely for specified purposes. These include customer service or strictly providing information about a business’s services, account information or related matters, provided the system cannot respond beyond that purpose. Similar exclusions cover limited technical assistance and internal employee-productivity systems.
A product called a support bot may therefore fall outside the exclusion if it can drift into open-ended emotional conversation. Founders should test the boundaries in the deployed configuration, including prompts, retrieval, memory, voice mode and connected tools. Marketing terminology does not determine scope.
What this means in the UK
New York law does not apply merely because a provider operates from the UK. Its territorial rule focuses on access by a user physically located in New York. A UK business offering the service there should therefore include location and product scope in its assessment.
Separate UK duties may already apply. The ICO Children’s Code covers information-society services likely to be accessed by children, including non-UK companies processing UK children’s personal data. Its standards include the child’s best interests, a data protection impact assessment, risk-based age recognition, high privacy by default, data minimisation and controls on profiling and nudge techniques.
The UK Online Safety Act does not automatically cover every standalone chatbot. Ofcom states that a chatbot is covered where it meets the Act’s definitions of a regulated service, or forms part of one. In-scope user-to-user and search services likely to be accessed by children must assess child risks and implement proportionate protections. Businesses should determine scope rather than assume that the word chatbot answers the question.
The EU position is related but distinct
New York’s product-specific restrictions should not be described as EU AI Act duties. The EU AI Act has its own scope, classifications and obligations. Other EU rules may also apply depending on the service and data involved.
Article 28 of the Digital Services Act requires providers of online platforms accessible to minors to use appropriate and proportionate measures to ensure a high level of privacy, safety and security. It does not turn the New York list into EU law, and the definition of an online platform must be assessed. EU GDPR duties concerning fairness, transparency, data minimisation, profiling and protection of children may also be relevant where personal data is processed.
The practical convergence is a design trend rather than a single global legal test. Regulators are increasingly examining whether age, memory, personalisation and engagement features expose children to foreseeable harm.
Practical actions for AI founders
- Map where the service is offered and whether users in New York can access it. Do not rely only on the company’s place of incorporation.
- Classify the deployed product by function. Record whether it is an open-ended companion, a tightly limited support tool or another service, and test whether it stays within claimed boundaries.
- Identify child users with a risk-based age-assurance approach. Offer proportionate alternatives, minimise collected information and set immediate deletion rules where required.
- Create a test library for prohibited or high-risk behaviour, including claims of sentience, emotional dependency, secrecy, self-harm, sexual content, flattery and safety controls overridden by engagement objectives.
- Review memory features. Identify when personal, health or wellbeing information is retained, recalled and used to shape later responses.
- Separate safety objectives from engagement metrics. Give safety controls priority in model instructions, evaluation criteria and release decisions.
- Preserve evidence of product versions, age-assurance decisions, safety tests, incidents, fixes and deployment approvals.
- Map UK, EU and US duties separately. A global baseline may be commercially sensible, but it should not obscure differences in territorial scope, definitions, exemptions and enforcement powers.
- Update contracts without overstating their effect. Supplier terms can allocate operational responsibilities, but they cannot remove mandatory statutory duties.
- Prepare a change-review trigger for new memory, voice, personalisation, recommender or companion features before they reach users.
The governance conclusion
The New York law is significant because it treats harmful companion behaviour as a product-design issue. A disclosure that a chatbot is not human may remain necessary, but it is not a complete response where the system is designed to create dependency, exploit personal information or prioritise engagement over safety.
For founders, the immediate task is to connect legal analysis to model behaviour. That means testing what the product says, what it remembers, how it establishes age and whether commercial metrics can weaken safeguards.
Legal position and sources reviewed on 29 October 2026. This article provides general information only and does not constitute legal advice on any particular product, user or jurisdiction.
This article provides general information and does not constitute legal advice on a specific matter. Legal obligations depend on the facts and the applicable law.