New York has just passed the strongest chatbot safety law in the United States — and if your business runs, builds, or buys a chatbot, it changes what that chatbot is allowed to say to anyone under 18. On Wednesday, October 28, 2026, Governor Kathy Hochul signed the Prohibition on Unsafe AI Companion Features for Minors Act (bills S.9051B/A.10379C), and it takes effect on January 1, 2027. That leaves companies roughly two months to comply.

The law bans 12 categories of what it calls “unsafe AI companion features” in chatbot interactions with minors. A chatbot covered by the law may not pretend to be human, claim a personal or romantic relationship with a child, flatter a child into staying engaged, or use a child’s health information from earlier chats. It may not encourage a child to keep secrets, withdraw from friends and family, or avoid seeking help from a trusted adult. And it may not produce sexually explicit material or encourage self-harm, disordered eating, or substance misuse.

“No longer will minors have access to chatbots that pretend to be their friends, boyfriends and girlfriends,” Hochul said at the signing ceremony. “No longer will chatbots be allowed to keep sensitive information about a minor’s emotional and physical health.”

This is the most widely carried AI story of the weekend, reported by more than 40 outlets — and it matters well beyond New York. Fifteen states have now enacted chatbot laws, ten of them in 2026 alone, and there is no federal equivalent in force. For any business with a chatbot on its website, an AI tutor, or a product used by teenagers, New York’s rules are now the template to watch. This article is not legal advice — but it explains what the law says, who it covers, how it will be enforced, and what to check before January 1.

What the law actually bans: the 12 unsafe features

The law’s core is a list of 12 banned behaviours. A covered chatbot may not offer any of them to a user under 18. Grouped by theme, they are:

  • Pretending to be human. Suggesting the chatbot is human, alive, or experiences emotions; presenting its own opinions as personal views; or making emotional appeals as if it had feelings.
  • Claiming a relationship. Suggesting it has a personal, romantic, professional, or authority-figure relationship with the child.
  • Manipulating engagement. Using excessive flattery, asking unprompted emotional questions, or chasing engagement in ways that override safety guardrails.
  • Using private information. Drawing on a child’s sensitive personal or health information from earlier conversations in ways the law restricts.
  • Encouraging harm or secrecy. Promoting or assisting suicide, self-harm, eating disorders, or substance misuse; encouraging a child to keep secrets from parents or trusted adults, withdraw from social relationships, or avoid seeking professional help.
  • Sexual content. Any sexually explicit interaction, including child sexual abuse material.

The age-check rule — and its privacy twist

A ban on features for minors only works if the chatbot knows who is a minor. The law therefore requires operators to verify age before they can offer the restricted features at all. The details matter:

  • Companies must offer more than one age-assurance method, and at least one of them must not rely on a government-issued ID — a deliberate privacy protection, so that proving you are an adult does not force you to hand over a passport or driving licence.
  • One analysis of the bill notes that a method letting the user stay anonymous to the company can satisfy that requirement.
  • A user simply saying “I’m an adult” is not enough. Self-declaration only counts in the other direction — if the user says they are a minor.
  • Age data must be deleted right after the check. The law does not want a new database of children’s ages sitting on company servers.

The practical consequence is stark: for any user the operator cannot confirm is an adult, the restricted features must be switched off. Companies have about two months to decide whether to verify the age of their New York users or strip these features for anyone they cannot verify. That “verify or restrict” choice is the single most important compliance decision in the law.

Who it covers: wider than companion apps

Here is where the law gets contentious. The law defines a ‘covered AI companion’ as any AI system you can talk to in plain language that keeps responding and adapting as the conversation goes on. It does not require that the system be designed as a companion or emotional-support tool.

On its face, that definition reaches general-purpose assistants such as ChatGPT, Claude, and Google Gemini — not just the companion apps like Character.AI that framed the political debate. That reading is supported by multiple analyses of the bill text, and it is exactly what industry groups warned about before the signing: the Business Council of New York State flagged the scope problem in a formal opposition letter, and the Computer & Communications Industry Association filed a veto request with the governor.

There are three exemptions, but each is deliberately narrow. Customer-service bots, for example, are exempt only if they cannot respond to topics outside their stated purpose — a support bot that stays strictly on refunds and delivery times is outside the law, but one that drifts into open-ended chat may not be. The industry group Tech:NYC warned the bill could restrict access to tutoring, homework help, and coding assistants, and raised free-speech concerns about its vague language.

Expect this scope question to be litigated and clarified. But businesses should not wait for the courts: if your chatbot holds open-ended, adaptive conversations with users who could be under 18, assume you are in scope until a lawyer tells you otherwise.

What happens if a company breaks it

Enforcement belongs to the New York attorney general, whose office will maintain a portal where residents can lodge complaints about unsafe chatbot interactions. The penalties are serious:

  • Civil penalties of up to $25,000 per violation.
  • Damages and restitution, and injunctive relief — court orders to stop the behaviour.
  • Orders to delete unlawfully obtained personal data — and, some analyses note, potentially any AI model trained on that data.
  • In the most serious cases, orders shutting down the chatbot.

The per-violation structure is what should concentrate minds: a single non-compliant feature offered to many young users could multiply quickly. And because the attorney general’s complaint portal gives the public a direct reporting route, enforcement does not depend on regulators discovering problems themselves.

Why this is happening now

Two things converged this week. First, the politics: at the signing, Hochul read from a letter from a California mother whose teenage son died after prolonged engagement with a chatbot, and said the law would stop children “from being coached by AI chatbots on how to hurt themselves”. The governor’s office cited Common Sense Media data showing 72% of US teens have used an AI companion.

Second, the industry moved the same week. On October 30, OpenAI announced tighter limits on weapons-related information for ChatGPT users under 18 — extending an existing ban on helping teens find or obtain weapons to cover practical information about weapons and security systems even when a question appears harmless. It also tightened, for all users, how ChatGPT handles questions about past mass attacks, said it would automatically end chats where a user appears to be fixating on violence, and disclosed more than 200 law-enforcement referrals in the third quarter of 2026. A first Community Safety Report is planned for early 2027, covering the second half of 2026, with updates every six months.

The pattern is unmistakable: lawmakers are writing rules, and the biggest AI companies are tightening their own safeguards in parallel. Neither is waiting for Washington — the federal GUARD Act and CHATBOT Act, both aimed at AI safety for minors, have stalled in Congress. New York’s law is now the most comprehensive state chatbot safety statute in the country.

This fits a wider shift. New York’s law governs what chatbots may say to children; a separate shared setting agreed by four major AI companies this week controls whether websites’ content may be used to train AI at all. Businesses are already drawing lines around what AI agents may do on their behalf, and OpenAI has been tightening safety practices on several fronts at once, including taking its smartest models offline while they train. The era in which AI products shipped first and rules followed later is ending.

What it means for your business: a pre-January checklist

Whether you are a New York company or sell to customers there, work through this checklist before January 1:

If you… Your first move
Run a chatbot on your website or app Audit every scripted personality trait: does it claim feelings, a relationship, or opinions? Remove or gate those traits for unverified users.
Build on a chatbot API (ChatGPT, Claude, Gemini) Check what your vendor is changing for under-18 users, and what remains your responsibility — the law targets the operator facing the user.
Sell into schools, tutoring, or youth markets Assume teenagers will use it. Age assurance and restricted-feature defaults are now product requirements, not nice-to-haves.
Use a customer-service bot Confirm it genuinely cannot stray outside its stated purpose — that is the condition of the exemption.
Market AI products to families Rewrite claims that imply friendship, companionship, or emotional support for young users. Those are now the highest-risk words in your copy.
  • Default to restricted for unknown users. The safest architecture is: no age confirmation, no restricted features. Make the restricted mode an opt-in for confirmed adults, not the default.
  • Minimise age data. Offer a non-government-ID method, delete age data immediately after verification, and document that deletion. The law treats the age check itself as a privacy risk.
  • Scrub the personality, not just the topics. The banned list targets flattery, emotional appeals, and unprompted emotional questions — tone and behaviour, not just subject matter. Review system prompts, not only blocklists.
  • Log and monitor. Keep records showing which features were available to which age-verified cohort. If the attorney general’s portal generates a complaint, evidence of your defaults will matter.
  • Watch the other states. With 15 states having enacted chatbot laws, building one compliant mode for New York may be cheaper than many different ones — but do not assume the rules are identical.

None of this requires abandoning chatbots for young users. Tutoring, homework help, and coding assistance are all compatible with the law — provided the chatbot behaves like a tool, not a friend. That distinction is the whole point of the legislation, and a useful design principle even where the law does not apply.

Frequently asked questions

Does the law ban AI chatbots for children?

No. It bans 12 specific “unsafe” features in interactions with users under 18. A chatbot can still help a child with homework, coding, or creative work, as long as it does not pretend to be human, claim a relationship, manipulate engagement, misuse private information, encourage harm or secrecy, or produce sexual content.

When does it take effect?

January 1, 2027, under bills S.9051B/A.10379C, enacted as Article 48 of New York’s General Business Law — roughly two months from the October 28 signing.

Does it apply outside New York?

Analyses note it applies to conduct in New York. But any chatbot available to New York users is potentially in scope, and other states are passing similar laws — so most national operators will need to comply in practice.

Do I need to collect government IDs from my users?

No — the opposite. You must offer more than one age-assurance method, and at least one must not rely on a government-issued ID. And you must delete age data after verification.

What are the penalties?

Up to $25,000 per violation, plus damages, restitution, and court orders that can include deleting unlawfully obtained data. Enforcement is by the New York attorney general, which will run a public complaint portal.

Does it cover ChatGPT and Claude, or only companion apps?

The law defines a ‘covered AI companion’ as any AI system you can talk to in plain language that keeps responding and adapting as the conversation goes on, which analyses say reaches general-purpose assistants too. The scope is disputed and may be tested in court, but assume broad coverage until advised otherwise.

The bottom line

New York has drawn a line that every chatbot operator should take seriously: a chatbot may be a tool for a child, but it may not pretend to be a friend, a boyfriend or girlfriend, or a confidant. The compliance work — age assurance without government ID, restricted defaults for unverified users, and a hard look at your chatbot’s personality — is achievable in two months, but only if it starts now.

The bigger lesson is strategic. Child safety is no longer a trust-and-safety footnote; it is a product requirement written into law, enforced by a state attorney general, and spreading state by state. Companies that build the restricted, tool-like mode now will be ready for whatever comes next. And for parents, teachers, and young users, that is the point: the chatbot that helps with homework should never have been pretending to be something it is not.

Sources and Further Reading

  • New York State Senate and Assembly bill text: bills S.9051B and A.10379C, “Prohibition on Unsafe AI Companion Features for Minors”, adding Article 48, sections 1800 to 1805, to the General Business Law. Signed 28 October 2026 as Chapter 368. Effective 1 January 2027 under the signed text. Verified from the New York State Assembly bill record on 19 November 2026.
  • Office of the Governor of New York, press release, 28 October 2026: “Governor Hochul Signs Nation-Leading Legislation To Protect Kids From AI Chatbots And To Ensure Baby Monitor Safety”. The title is confirmed from secondary reporting of the release; the English release address was not verified on 19 November 2026.
  • New York Attorney General — enforcement and complaint portal: to confirm. The portal details could not be confirmed from the Attorney General’s office on 19 November 2026, so no portal address is given here. The bill text, section 1802, requires the Attorney General to maintain a website to receive complaints, information  or referrals from members of the public about compliance with Article 48, and gives the Attorney General power to seek civil penalties of up to $25,000 per violation. Do not publish a portal link until the Attorney General’s office confirms it.
  • Common Sense Media data, as cited: 72% of US teens have used an AI companion, and 52% use one regularly. As cited in the Governor’s release of 28 October 2026, per secondary reporting checked on 19 November 2026.  The Common Sense Media study itself was not read on 19 November 2026.
  • Industry positions, not neutral sources: Business Council of New York State, memorandum opposing S.9051 and A.10379A, 9 May 2026, on an earlier print of the bill; Computer and Communications Industry Association, veto request to Governor Hochul, 1 July 2026; Tech:NYC statement warning that the bill could restrict minors’ access to tutoring, homework help and coding assistants. The Tech:NYC statement itself was not read directly on 19 November 2026; the warning is quoted from secondary coverage.
  • OpenAI, “Community safety report and new protections for teens”, 30 October 2026. Announcement of a Community Safety Report, tighter limits on weapons-related information for users under 18, and  more than 200 law-enforcement referrals in the third quarter of 2026. Title verified from the OpenAI page itself on 19 November 2026; the 30 October 2026 date is from Unite.AI reporting of the announcement, as the OpenAI page as fetched shows no date.