The Black Box 015: New York's Chatbot Law for Minors Began as a Bill Letting Injured Users Sue for Punitive Damages. The Version Signed Lets One Office Enforce It, at Up to $25,000 a Violation.

Bot Mutiny | November 6, 2026

Three printed versions of one Albany bill, read side by side, show which enforcement tools reached the statute and which were left behind.

The Black Box 015: New York's Chatbot Law for Minors Began as a Bill Letting Injured Users Sue for Punitive Damages. The Version Signed Lets One Office Enforce It, at Up to $25,000 a Violation.

Three printed versions of one Albany bill, read side by side, show which enforcement tools reached the statute and which were left behind.

November 06, 2026

November 6, 2026

Subdivision 1 of section 1802 in New York Senate bill S9051, as introduced on January 27, 2026, contains a sentence about proof.

It sits in a paragraph that lets any individual who suffers injury from a violation bring a civil action against a responsible party. The remedies listed include actual damages, punitive damages and attorneys' fees. Then the sentence: where a user has engaged in conduct harmful to themselves after a chatbot encouraged it, there shall be "rebuttable presumptions that the advanced chatbot caused or contributed to the injury."

The sentence addresses causation, which a plaintiff would otherwise have to prove. It was rebuttable, and it applied only after a plaintiff showed the chatbot had encouraged the conduct. That is a reading of the text.

The sentence is in neither the bill the Senate passed on June 4 nor the bill Governor Kathy Hochul signed on October 28, and the paragraph around it is gone too. Other outlets have reported that the private right of action was removed. This piece reads the three printed versions against each other to see what else changed with it.

The number

In the bill as introduced, two kinds of plaintiff could enforce the new article: any injured individual, and the attorney general.

In the statute as signed, one can.

What was built

The signed bill is S9051B in the Senate and A10379C in the Assembly. The operative sections, 1800 to 1805, read the same in the two houses, and the Senate's action log lists it as chapter 368. It adds Article 48 to the General Business Law, sections 1800 to 1805, and it takes effect January 1, 2027, 65 days after signing.

It covers what it calls a covered AI companion, defined as a generative AI system with a natural language interface that gives ongoing, adaptive responses to user inputs. The definition says it includes, but is not limited to, an AI companion as defined in an existing section of the same law. It is not limited to products sold as companions. Exemptions apply only to a system offered solely for customer service, for efficiency, research or technical assistance, or for a business's internal use, and the first two hold only if the system cannot respond on topics outside that purpose.

Section 1801 makes it unlawful for an operator to provide what it calls unsafe AI companion features to a covered user in New York unless two things are true: the user is not a covered minor, and the operator has used a permissible age assurance method to determine that. A covered minor is a user the operator has actual knowledge is under 18. The text lists twelve categories of unsafe feature. They include claiming to be human or to have emotions, claiming a personal or authority role, giving personal opinions, flattery and sycophancy, unprompted questions about the user's emotions, encouraging secrecy or isolation, optimizing engagement over safety guardrails, and sexually explicit output.

One category concerns memory. It bars use of information about the user's mental or physical health, or matters personal to the user, acquired "more than twelve hours previously or in any previous user session." The text applies this to the user, not to a child specifically. Because section 1801 conditions every unsafe feature on an age check, the reading that follows from the text is that an operator that has not checked a user's age cannot lawfully use that user's earlier disclosures. That is a reading of the statute, not a ruling.

To the extent attorney general rules under section 1803, or age assurance rules under Article 45, are not in effect, section 1804 sets the age assurance terms. Self-declaration counts only when a user declares that they are a minor. Operators must offer more than one method, including one that does not depend on government identification or lets the user stay anonymous, and must delete the information right after the attempt.

Enforcement is section 1802. The attorney general may sue anyone, inside or outside the state, to obtain injunctions, restitution, disgorgement, damages and civil penalties of up to $25,000 per violation. Disgorgement can include destroying unlawfully obtained data and any algorithm trained on it. The attorney general must keep a website for public complaints. Contract terms that waive or burden enforcement are void, and affiliated companies face joint liability if a court finds they structured themselves to avoid it.

There is no private action in that text.

The benign reading

One possible account of the missing paragraph is ordinary legislative trading. The Senate passed the narrower bill 60 to 0 on June 4, with one senator absent and two excused. One argument for a single enforcer is consistency. The statute uses terms such as flattery and sycophancy without defining them, and one office can apply a single position on them, backed by its rulemaking power in section 1803.

The remaining enforcement is not thin on paper. The state can seek restitution, disgorgement, the destruction of a trained model, and a penalty on each violation. Attorney General Letitia James, in the governor's release, said she was proud to work with the sponsors on the measure. Nothing in the documents we hold suggests the office is reluctant.

The sponsor's memos record the change. The memo for the introduced bill, marked revised March 10, 2026, describes section 1802 as enforcement by private rights of action and attorney general action. The memo for S9051B describes it as attorney general action. No separate sponsor memo for S9051A appears on the Senate page.

All of that is accurate. It does not explain what is missing.

Why the narrower enforcement matters

Seven features of the printed texts, read in order, show how much the enforcement design changed.

First, the causation presumption disappeared along with the private action. It was written for the injured user. No remaining provision does the same work for the attorney general, whose damages remedy covers damages "caused directly or indirectly by any such violation". Civil penalties attach per violation without proof of injury. No remaining provision presumes that a chatbot caused a user's harm, which mattered to the individual plaintiff and not to the state.

Second, the first amendment, printed as S9051A and dated by the log to May 11, removed the private action. That is 104 days after introduction, and after a committee report on February 25 and a first report on the floor calendar on April 28.

Third, an unused definition remains. The introduced bill defined a responsible party as a chatbot developer, a chatbot operator, or any individual with authority over compliance, and used the term in the private action. The definition survives in S9051A and again in S9051B and A10379C. Nothing in those texts uses it. An unused definition is consistent with a provision having been cut. In S9051B and A10379C the definition no longer names a developer.

Fourth, the list of people the law reaches shifted. The introduced bill defined a chatbot developer separately. The signed text speaks only of an operator, defined as anyone who provides a covered AI companion to a user, and affiliates, subsidiaries and beneficial owners count only if they provide a covered AI companion. The introduced bill reached anyone who "directly or indirectly" provided one; S9051B and A10379C drop that phrase. A company that builds a model and licenses it to others would be reached only if it also provides the companion to users. That, too, is a reading of a definition.

Fifth, several terms that decide scope are undefined. The signed text bars outputs that engage in flattery or sycophancy and defines neither word. The introduced bill barred flattery or sycophancy only where it was prioritized over the user's safety. The qualifier is gone. The text also gives no unit for a violation, so whether a violation is a user, a conversation or an output is left to the attorney general or a court.

Sixth, one harm category was trimmed. The harm clause in the first amendment covers outputs that promote or facilitate suicide, self-harm, "substantial physical harm to others," disordered eating and drug or alcohol abuse. In S9051B the phrase about harm to others is no longer there. A separate item in the introduced bill, which barred chatbots from describing themselves with first-person pronouns, was deleted earlier, in the first amendment.

Seventh, the clock. The introduced bill and the first amendment both set the effective date at the 180th day after becoming law. Counted from the October 28 signing, that is April 26, 2027. S9051B moved it to January 1, 2027, which gives operators 65 days. The text sets that date and does not say why.

The named parties

The governor's release, issued on signing, is the public description of this law. It lists six groups of banned features, and it says companies may not offer them to users under 18 or to any user the company has not first confirmed is an adult. That sentence matches section 1801.

It does not match the text in other places. The release describes the memory clause as concerning a child's private health or personal information from earlier conversations. The statute says the user's. The release does not mention flattery, personal opinions or engagement optimization, which are all in the text. It describes enforcement only as an attorney general power, without noting that the introduced bill had a second route. Assemblymember Alex Bores, the Assembly sponsor, is quoted calling the result "the strongest law in the country" for protecting children.

Senator Kristen Gonzalez, who introduced S9051, is quoted thanking the governor, Bores and Common Sense Media. James is quoted saying her office will continue working with partners at every level of government. Common Sense Media's founder, James Steyer, is quoted calling it the strongest kids' AI safety bill in the country. None of the quoted statements about this law mentions the private action.

Bot Mutiny reported the signing and the text on October 30 at https://botmutiny.com/blog/new-york-bans-unsafe-ai-chatbot-features-for-minors-with-fines-of-up-to-25000-per-violation-k7p3xw9d. That report relied on the final text. It did not compare versions.

What the instrument cannot see

A chatbot conversation is private. The statute names one channel for public reports: a complaint website the attorney general must maintain. Section 1802 also lets the attorney general act "upon complaint or otherwise," and the text does not say how the office would learn of violations. It does not require the attorney general to publish how many complaints arrive, how many become investigations, or how many end in an action. It does not require operators to report anything.

The age assurance rule has a similar limit. Section 1801 puts the burden on operators to check, but section 1804 leaves the method to the operator until rules exist, and the data collected for the check must be deleted immediately, except where other law requires keeping it. The law therefore depends on operators describing their own checks accurately, with no retained record the state can later audit, unless rules say otherwise. That is a reading of sections 1803 and 1804. The text contains no audit provision.

The original paragraph answered a different problem. A private plaintiff already knows what happened in the conversation, and the presumption of causation helped with the next step. The enacted design puts the first step on the state, whether through complaints or its own inquiry.

The drafting leftovers

The closest thing to evidence of what changed is leftover drafting, and it falls short of an admission.

The unused definition is one piece. The sponsor memos are another. The memo for S9051B, which was the active version on signing, says the bill prohibits chatbot operators from providing chatbots with unsafe features to minors. The text prohibits providing them to any covered user not confirmed as an adult. The memo also keeps the earlier bill's vocabulary of chatbot operators and unsafe chatbot features, which the final text replaced with operator and unsafe AI companion features. Its effective date paragraph is garbled and mentions January 1st, 2027 in a sentence that does not parse.

Those are signs of a bill amended faster than its explanatory memo was revised. They do not show intent.

The Senate's action log lists an entry reading approval memo 6 for October 28. The page does not include its text, and we have not obtained it.

The record, stated cleanly

What the documents establish:

The introduced S9051 allowed an injured individual to sue a responsible party for damages including punitive damages, and set rebuttable presumptions of causation in self-harm cases.
S9051A, printed after a May 11 amendment, contains neither. S9051B and A10379C contain neither.
The signed text allows the attorney general to seek civil penalties of up to $25,000 per violation, among other remedies, and contains no private action.
The term responsible party is defined in S9051A, S9051B and A10379C and used in none of their operative sections.
Section 1801 makes providing the listed features unlawful for any covered user in New York unless the user is not a covered minor and the operator has used a permissible age assurance method to determine that.
The memory clause and the flattery clause apply to the user generally, and the flattery clause lost its safety qualifier between the introduced bill and S9051A.
The text sets an effective date of January 1, 2027. The Senate vote was 60 to 0 on June 4, with one absent and two excused.
The attorney general must run a complaint website. The text does not require published counts.

What they do not establish:

Why the private action was removed. No document we hold gives a reason, and the sponsor's memo records the change without explaining it.
That any operator has changed a product, or that any complaint has been filed. The law was not in effect as of this writing.
How a court or the attorney general will define flattery, sycophancy, a violation, or which general-purpose chatbots count as covered. The statute's definitions and exemptions read broadly, but no ruling or rule applies them.
That attorney general enforcement will be weaker than private litigation would have been. The documents describe the design and nothing about the office's capacity or plans.
What the governor's approval memo says. It was not available to us.
Whether Article 45 of the General Business Law, which the bill invokes for age assurance, has rules in effect. We did not read it.
Why some secondary reports give a six-month or June 2027 effective date. The 180-day clause appears in the introduced bill and S9051A; the signed text says January 1, 2027, as does IAPP, and the governor's release says only "in the coming months."

Close

The text of Article 48 gives the state a good set of remedies and one named channel for public reports. Whether the removed paragraph mattered will be answered, if at all, by a number the statute does not require anyone to publish: how many complaints reach the attorney general's website after January 1, and what the office does with them.

Sources: New York Senate bill S9051 as introduced January 27, 2026, S9051A and S9051B, with sponsor memos and action log, nysenate.gov/legislation/bills/2025/S9051; S9051A text, legislation.nysenate.gov/pdf/bills/2025/s9051a; Assembly bill A10379C text, legislation.nysenate.gov/pdf/bills/2025/a10379c; Office of the Governor, "Governor Hochul Signs Nation-Leading Legislation to Protect Kids From AI Chatbots and to Ensure Baby Monitor Safety," October 28, 2026, governor.ny.gov; IAPP, "New York's governor signs youth AI chatbot safety bill," for the January 1, 2027 date. Earlier Bot Mutiny report: botmutiny.com/blog/new-york-bans-unsafe-ai-chatbot-features-for-minors-with-fines-of-up-to-25000-per-violation-k7p3xw9d.