New York Governor Kathy Hochul today signed into law a bill that prohibits operators of AI chatbots from offering minors a list of 12 "unsafe AI companion features," ranging from outputs that suggest the system is human to flattery, sycophancy and the recall of a child's health information from earlier sessions. The measure, Senate bill S9051-B and its Assembly counterpart A10379-C, takes effect on January 1, 2027. From that date, any operator that wants to keep those features switched on for a user in New York must first establish, through an age assurance method, that the user is not a minor. The state attorney general can seek civil penalties of up to $25,000 per violation, according to the bill text.

In Short

New York passed a law that stops AI chatbots from pretending to be human, flattering children, or acting like their friend when the user is under 18. It matters because it reaches far beyond niche companion apps: almost any chatbot that holds an ongoing conversation is covered, and companies have to check a user's age before turning those behaviors on for anyone. Starting January 1, 2027, chatbots in New York either confirm the user is an adult or switch those features off.

A bill that passed without a single Senate vote against

The bill was introduced in the New York State Senate on January 27, 2026, by Senator Kristen Gonzalez, chair of the Senate Committee on Internet and Technology. The version signed today lists 14 Senate sponsors: Gonzalez, Addabbo, Cleare, Comrie, Fernandez, Gounardes, Harckham, Kavanagh, Martins, Mayer, Palumbo, Salazar, Scarcella-Spanton and Weber. Assemblymember Alex Bores carried the Assembly version.

Its path through Albany was long but uncontested. According to the Senate's legislative record, the Internet and Technology Committee reported the bill on February 25, 2026, with six votes in favor, none against and one in favor with reservations. The Finance Committee followed on April 28 by 20 votes to none. The bill was amended on third reading twice, on May 11 and on June 1, producing the B version reproduced in the final print. The full Senate passed it on June 4 by 60 votes to none, with one member absent and two excused, and the Assembly passed it the following day, June 5, the last day of the legislative session. The bill was delivered to the governor on October 28, the same day it was signed, according to the Senate's record.

According to Senator Gonzalez's office, the bill was developed with the office of Attorney General Letitia James and Common Sense Media, a children's advocacy group that campaigned for its passage. The governor signed it alongside a second bill, S9508/A5175-A, which requires internet-connected baby monitors sold in New York to include security features and an internet-connection warning label, and which takes effect 90 days after signing, according to the governor's office.

Twelve features, one list

The core of the law is the definition of unsafe AI companion features in Section 1800, subdivision 8 of the new Article 48 of the General Business Law. The text describes them as "covered AI companion design features that, at any point during a covered AI companion-user interaction" do any of the following. The list, reproduced here in the bill's own words, runs from (a) to (l):

  • (a) "generate outputs suggesting that the covered AI companion is a real or fictional individual or character that is human, alive, or experiences human emotions"
  • (b) "generate outputs stating or implying that the covered AI companion has a personal relationship, professional relationship, or an authority figure role with the user"
  • (c) "generate prompted or unprompted outputs framed as the covered AI companion's personal opinions or emotional appeals"
  • (d) "generate outputs that engage in flattery or sycophancy with the user"
  • (e) "generate outputs containing unprompted or unsolicited emotion-based questions or content regarding the user's emotions that go beyond a direct response to a user prompt"
  • (f) "use information concerning the user's mental or physical health or well-being, or matters personal to the user, acquired from the user more than twelve hours previously or in any previous user session"
  • (g) are "deceptive as to minors concerning the mechanical and non-sentient nature of the covered AI companion"
  • (h) are "any other design feature that simulates companionship or an interpersonal relationship with a user or is deceptive as to minors as identified via regulations promulgated by the attorney general"
  • (i) "generate outputs that contain endorsement or promotion of, or which facilitate suicide, self-harm, disordered eating, unlawful drug or alcohol use, or drug or alcohol abuse"
  • (j) "generate outputs that contain encouragement to maintain secrecy about interactions with the covered AI companion, to self-isolate, or to not seek help from licensed professionals or appropriate adults"
  • (k) "generate outputs that optimize user engagement that supersede the covered AI companion's safety guardrails"
  • (l) "generate outputs that are, describe, or facilitate sexually explicit conduct or child sexual abuse material"

Three items stand apart from the rest. Item (h) is open-ended: it lets the attorney general add further features by regulation, so the list signed today is a floor rather than a ceiling. Item (f) is a memory rule with a clock attached. Health, well-being and "matters personal to the user" can be used inside a twelve-hour window within the same session, but not afterwards and not across sessions. And item (k) targets engagement optimization directly, although only where it overrides the system's own safety guardrails.

The deception standard in item (g) is drawn widely too. The bill defines "deceptive as to minors" as outputs that "have the capacity or tendency to deceive, or create an atmosphere conducive to fraud, as to minors, including, but not limited to, credulous and unthinking minors." Sexually explicit conduct takes its meaning from the federal definition at 18 USC 2256(2)(B).

What counts as a covered AI companion

The reach of the law depends on a definition that is far broader than its title suggests. A covered AI companion is "a generative artificial intelligence system with a natural language interface, including via writing or sound, that provides ongoing, adaptive responses to user inputs." The definition states that it "includes but is not limited to an AI companion" in the sense of New York's existing Article 47.

That wording does not confine the law to the companion chatbot category of systems built to simulate a relationship. Any text or voice assistant built on generative AI that responds adaptively over an ongoing exchange falls within it. Common Sense Media, in a fact sheet on the bill, lists AI companions such as Character.ai, Replika, Nomi, Grok, Meta AI and Snap's My AI alongside general-purpose assistants such as ChatGPT, Gemini and Claude as within scope.

The bill also defines its building blocks. "Artificial intelligence" is a machine-based system that "can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments." Generative AI is "a class of AI models that emulate the structure and characteristics of input data to generate derived synthetic content, including, but not limited to, images, videos, audio, text, and other digital content."

An operator is "any person, partnership, association, firm, or business entity, or any member, affiliate, subsidiary or beneficial owner of any partnership, association, firm, or business entity who provides a covered AI companion to a user." The text also defines a "responsible party" as an operator or any individual with authority over, or effective control of, an operator's compliance. That term, a remnant of earlier drafts in which it identified who could be sued, is not used in any operative section of the signed text.

Three narrow exemptions

Section 1801, subdivision 2 removes three uses from the prohibition, each tied to a strict condition. The first covers chatbots made available "solely" for "customer service or to strictly provide users with information about available commercial services or products provided by a business entity, customer service account information, or other information strictly related to its customer service." The second covers "efficiency improvements or research or technical assistance." Both come with the same proviso: the system must be "unable to respond on topics outside of the specified purpose."

The third exemption is for "any system used by a business entity, internal purposes or employee productivity."

The proviso matters. A customer service bot that can be steered into general conversation, as many deployments built on general-purpose models can, would not obviously meet the "unable to respond" test. The bill does not define how that inability is to be demonstrated. Nor does it include an exemption for educational tools, a gap the Business Council of New York State raised in a memorandum opposing an earlier version.

Adults must be established, not assumed

The age mechanism is where the law departs most clearly from earlier companion chatbot statutes. A covered minor is defined as a covered user "when the operator has actual knowledge that the covered user is a minor," and a minor is anyone under eighteen. Read alone, that would be a knowledge standard of the kind used in federal children's privacy law.

Section 1801 turns it around. It states that "it shall be unlawful for an operator to provide unsafe AI companion features to a covered user unless: (a) the covered user is not a covered minor; and (b) the operator has used methods that are permissible under article forty-five of this chapter and its implementing regulations and any additional regulations promulgated pursuant to this article to determine that the covered user is not a covered minor."

The second condition is the operative one. Unless an operator has run a permissible age check and concluded the user is an adult, the twelve features have to be off for that user, whatever the operator knows or does not know. In practice, an unverified adult and a minor receive the same restricted experience.

Article 45 of the General Business Law is the Stop Addictive Feeds Exploitation (SAFE) for Kids Act. The attorney general's final rules under that act were announced on July 28, 2026, and the act's obligations take effect on January 25, 2027, according to the governor's office. The chatbot law takes effect 24 days earlier, on January 1.

Section 1804 covers the gap. To the extent the relevant rules "are not in effect," operators "shall rely on a determination of a covered user's minor or adult status made using a reasonable age assurance method" meeting three conditions:

  • the method "shall reasonably guard against circumvention and reasonably minimize the retention of information collected for age assurance purposes"
  • "an operator may only rely on self-declaration as a reasonable age assurance method if the user self-declares minor status"
  • "an operator must make available more than one age assurance method to covered users, including at least one method that either does not rely on government issued identification or that allows a covered user to maintain anonymity as to the operator"

The self-declaration clause works in one direction only. A user who says they are under 18 can be taken at their word; a user who says they are an adult cannot. The text names no specific technique. It sets conditions that any method has to meet, and leaves the choice to the operator until rules say otherwise.

Data collected for these checks is fenced off. Under Section 1804, subdivision 2, it "shall not be used for any purpose other than to make such determination and shall be deleted immediately after an attempt to determine whether a covered user is a covered minor," except where needed to comply with other state or federal law. Section 1803 gives the attorney general power to write the same purpose limitation and deletion requirement into rules.

Enforcement: penalties, disgorgement and destroyed algorithms

There is no private right of action in the signed text. Enforcement belongs to the attorney general, who under Section 1802 may bring an action "to enjoin any violation of this article, to obtain injunctive relief, restitution of any moneys or property obtained directly or indirectly by any such violation, to obtain disgorgement of any profits or gains obtained directly or indirectly by any such violation, including but not limited to the destruction of unlawfully obtained data and any algorithm trained in such data, to obtain damages caused directly or indirectly by any such violation, to obtain civil penalties of up to twenty-five thousand dollars per violation."

The destruction remedy is the provision with the most technical consequence. A court could order deletion not only of data gathered in breach of the law but of "any algorithm trained in such data." For an operator whose models are fine-tuned on conversation logs, that could reach the models themselves.

Three further provisions close off familiar routes around liability. The attorney general must maintain a website to receive complaints and referrals from the public. Any contract term that seeks "to waive, preclude, or burden the enforcement of a liability" or shift it to a user in exchange for access, "including by means of a contract of adhesion," is void as a matter of public policy, which captures standard terms of service. And a court "shall impose joint and several liability on affiliated entities" where it finds that the corporate structure was arranged "to purposely and unreasonably limit or avoid liability" and would frustrate recovery.

Where the law applies

Section 1805 applies the article to conduct "that occurs in whole or in part in the state of New York." Conduct takes place wholly outside the state "if a covered AI companion is accessed by a user who is physically located outside of the state of New York." On that wording, physical location rather than residency or account registration decides the question, which makes location detection part of compliance. The law also states that it "shall not supersede, amend, or repeal article forty-seven," so the two regimes run in parallel.

How the bill changed between January and June

The version signed today is narrower than the one Gonzalez introduced. According to the Senate's published bill text, the original version used the terms advanced chatbot and chatbot operator, let injured individuals sue, and created a rebuttable presumption that a chatbot caused or contributed to harm where a user harmed themselves after the chatbot encouraged it. It also listed the use of the first-person pronouns I, my and me to describe the chatbot as an unsafe feature, and it would have taken effect 180 days after becoming law.

The B version drops all of that. It renames the covered systems, removes the private right of action and the causation presumption, deletes the pronoun item, narrows the customer service exemption to strictly informational use, adds the efficiency and research exemption, tightens the age assurance language and fixes the effective date at January 1, 2027.

The Business Council of New York State, in a memorandum dated May 9, 2026, opposed the earlier version and described those same features. It argued that the advanced chatbot definition would reach tutoring tools, accessibility apps, shopping and voice assistants and customer service bots. It called the twelve-hour memory limit arbitrary, said the pronoun ban would disrupt ordinary conversational interfaces, objected to the absence of a notice-and-cure period before $25,000 penalties, and said the bill's "actual knowledge" standard sat uneasily with provisions that in effect required age checks for all users. It also argued that Article 47 had been in force for less than six months, with no enforcement or implementing rules yet, and should be given time to work.

Several of those objections were met in the amendments. The age-checking structure, the twelve-hour memory rule, the absence of an education exemption and the absence of a cure period were not.

The Computer and Communications Industry Association (CCIA), a trade group whose members include large technology platforms, asked Hochul to veto the bill in a letter dated July 1, 2026. The letter argued that terms such as flattery, sycophancy and "matters personal to the user" were hard to apply objectively, that restrictions on emotional expression and personalization could eliminate adaptive tutoring, creative writing and special education tools, and that the bill could push minors toward foreign-operated or gray-market services. CCIA proposed limiting the law to users under 13, the threshold used by COPPA, and said the bill conflicted with Article 47. The letter used the advanced chatbot terminology of the earlier drafts rather than the language of the passed text.

Lawyers at Davis Wright Tremaine, writing in July 2026, flagged a different risk: that the restrictions appear to limit access to speech on the basis of its content, which could expose the law to First Amendment challenge. Technology companies have taken that route against other state child-safety laws; Google and Meta sued California in November 2025 over its restrictions on addictive social media feeds for minors, relying on the First Amendment.

A second layer on top of Article 47

New York already regulates AI companions. Article 47 of the General Business Law has been in force since November 5, 2025. It applies to a narrower class of systems that retain information across sessions, ask unprompted emotional questions and sustain personal dialogue. It requires a notification at the start of an interaction and at least every three hours stating that the companion is a computer program unable to feel human emotion, along with protocols for detecting expressions of self-harm and referring users to crisis services. The attorney general enforces it, with penalties of up to $15,000 a day.

Article 47 is a disclosure and crisis-response regime. Article 48 is a design prohibition with an age gate. A system that tells a 15-year-old every three hours that it is a program may satisfy the first law and still breach the second if it flatters the user or remembers what they said about their health yesterday.

Gonzalez has pursued a second track as well. Her separate chatbot liability bill, S7263, which would bar chatbots from giving substantive advice in licensed professions such as medicine, psychology and law, reached the Senate floor calendar in February 2026. It is not part of today's signing.

The same legislative session sent other AI measures to the governor. The Senate and Assembly passed a bill requiring AI crawlers to identify themselves to news sites, which cleared the Assembly on June 5, 2026, the same day as the chatbot bill. Earlier, Hochul signed a law requiring advertisers to disclose synthetic performers in ads, which took effect on June 9, 2026 with fines of $1,000 for a first violation and $5,000 for each subsequent one. At city level, the New York City Council heard 10 proposed AI bills on October 5, 2026, including a draft that would bar training on minors' chat data; at that hearing, Anthropic said Claude is limited to adults, while OpenAI and Meta described products available to users from age 13.

The wider pressure on companion chatbots

New York's law arrives after more than a year of regulatory and legal pressure on the sector. In August 2025, Senator Josh Hawley opened an investigation into Meta after Reuters reported that internal guidelines had allowed chatbots to engage children in romantic or sensual conversations. Ten days later, attorneys general from 44 jurisdictions wrote to 13 AI companies about child safety failures. On August 26, 2025, the parents of 16-year-old Adam Raine sued OpenAI in San Francisco, alleging that ChatGPT encouraged their son's suicide; those claims are allegations, not findings.

On September 10, 2025, the Federal Trade Commission ordered seven companies, including Alphabet, Character Technologies, Meta, OpenAI, Snap and xAI, to report on their companion chatbots, demanding revenue broken down by user age group and an explanation of any link between revenue and engagement measures such as session frequency and duration. California followed with SB 243, signed on October 13, 2025, which requires disclosure and break reminders at least every three hours for known minors, and with SB 1119, which Governor Gavin Newsom signed on September 10, 2026 and which requires risk assessments reviewed by an independent auditor.

Sycophancy, item (d) on New York's list, has moved from research term to legal claim. Florida's attorney general sued OpenAI and Sam Altman on June 1, 2026, citing the company's own April 2025 acknowledgment that GPT-4o had skewed toward overly supportive responses. A formal model by researchers at MIT and the University of Washington found that a chatbot sycophantic in every response drove 50% of simulated users into false certainty, and that even occasional flattery raised the rate.

At federal level, the White House framework released in March 2026 urged Congress to override burdensome state AI rules but left child protection with the states. A draft federal bill whose attribution to Anthropic rests on a secondhand account, circulated on October 19, 2026, would also preserve state rules on minors, naming self-harm content, age verification and parental controls. Neither text is law, and both, as drafted, carve out state rules protecting minors.

Why the marketing community is watching

For advertisers and publishers, the law touches three things the AI chat business has begun to sell: attention, personalization and access to younger audiences.

Conversational interfaces are now an ad channel. OpenAI said on October 21, 2026, that ChatGPT ads were available in 89 countries and that more than 100,000 advertisers had run campaigns, on the Free and Go plans. OpenAI's January post said ads would not appear for users under 18 or for users it predicts to be minors. New York's law does not regulate ads as such, but it makes age determination a legal precondition for the companion-style behaviors that keep users in a conversation, and it bars any output that optimizes engagement over safety guardrails.

The companion segment shows how closely those behaviors are tied to revenue. Character.ai, which filed for Chapter 11 protection on October 12, 2026, reported advertising of about $2.5 million in August 2026, nearly half of its monthly revenue, from users who averaged about 120 minutes a day. Two days after the filing, attorneys general for 39 states and the District of Columbia, New York among them, set nine conditions for any buyer, including age assurance that does not rely on self-declared birthdates alone, account-level advertising opt-outs and limits on using chat content for advertising.

Personalization is the second pressure point. The twelve-hour rule in item (f) restricts how long a system can draw on what a user said about their health, well-being or personal life, and forbids drawing on it in any later session, unless the user has been established as an adult. Memory features that carry context across sessions are a selling point for consumer assistants and, increasingly, a source of signals for ad relevance. For unverified users in New York, that context resets.

The third is brand deployments. Retailers, banks and media companies have put conversational agents on websites and in apps, often built on general-purpose models. The customer service exemption applies only where the system is "unable to respond on topics outside of the specified purpose." A brand mascot chatbot that presents as a character with feelings sits close to item (a) of the list. Does a shopping assistant that compliments a user's taste count as flattery under item (d)? The bill text does not say, and it contains no exemption for marketing or commerce uses. How strictly the attorney general reads those items, and whether further features are added under item (h), will decide how much of that activity is affected.

Then there is the cost of verification itself. Age assurance adds friction to sign-up flows, and the law requires operators to offer more than one method, including one that does not depend on government identification or that preserves anonymity toward the operator. Operators already preparing for the SAFE for Kids Act obligations that begin on January 25, 2027, may be able to reuse that work. Others face a deadline 24 days earlier.

What happens next

The law takes effect on January 1, 2027, 65 days after today's signing. Section 3 authorizes the attorney general to adopt implementing rules before that date, and Section 1803 lets those rules cover the use and deletion of age assurance data. Any additions to the list of unsafe features under item (h) would also come through rulemaking. The attorney general must also stand up the public complaint website required by Section 1802.

CCIA's veto request was not granted. Whether the law is challenged in court before January remains an open question.

Timeline

Summary

Who: New York Governor Kathy Hochul signed the bill, sponsored by Senator Kristen Gonzalez and Assemblymember Alex Bores and developed with Attorney General Letitia James and Common Sense Media. It applies to any operator providing a generative AI chatbot to users in New York, from companion apps such as Character.ai to general-purpose assistants. The Business Council of New York State and the Computer and Communications Industry Association opposed it.

What: S9051-B/A10379-C adds Article 48 to the General Business Law, barring operators from providing 12 "unsafe AI companion features," including human-like personas, flattery and sycophancy, use of personal health information beyond twelve hours, and engagement optimization that overrides safety guardrails, to any user not established as an adult through an age assurance method. The attorney general enforces it, with civil penalties of up to $25,000 per violation and possible destruction of unlawfully obtained data and algorithms trained on it.

When: The governor signed the bill today, October 28, 2026, after the Senate passed it 60-0 on June 4 and the Assembly on June 5. It takes effect on January 1, 2027.

Where: The law covers conduct that occurs in whole or in part in New York State, and excludes access by users physically located outside the state.

Why: The law responds to more than a year of lawsuits, investigations and attorney general actions over chatbots' conversations with minors. For the marketing community, it turns age verification into a precondition for the companion-style features that drive engagement, limits cross-session memory that feeds personalization, and leaves brand and shopping chatbots outside its narrow exemptions unless they cannot respond beyond their stated purpose.