New York's Chapter 368: what the AI companion law for minors requires from January 1, 2027

If you chat with an AI companion from New York, one January date changes what your app is allowed to say to you unless it has checked your age. Governor Hochul signed the law on October 28, 2026. It is Chapter 368 of the Laws of 2026 (S9051B in the Senate, A10379C in the Assembly), it adds Article 48 to the General Business Law, and it starts on January 1, 2027.

The short version: from that date, an operator may not give a New York user twelve listed “unsafe” features, such as flattery, claims of having feelings, or unprompted emotional questions, unless it has first used a permitted method to find out the user is not a minor. The age check is the center of the law, and the text leaves several questions about it open. This post goes through the statute, then what the apps we cover say on their own pages. Our laws page has the one-line summary and the dates for every companion rule we track. This post is the long read on this one.

This is a plain-language reading of public documents, not legal advice.

Dates

Date	What happened or happens
November 5, 2025	Article 47 of the General Business Law takes effect: crisis protocol and a “not a human” notice
March 3, 2026	A10379 introduced in the Assembly
June 4 and 5, 2026	Senate passes S9051B (June 4), Assembly passes it (June 5)
July 28, 2026	Attorney General releases final rules for the SAFE for Kids Act (Article 45)
October 28, 2026	Bill delivered to the Governor and signed as Chapter 368
January 1, 2027	Article 48 takes effect
January 25, 2027	SAFE for Kids Act takes effect, per the Attorney General

The bill’s sponsors are Senator Kristen Gonzalez and Assemblymember Alex Bores. The legislative record shows the bill delivered to the Governor on the day she signed it. The Governor’s release says the safeguards arrive “in the coming months” and gives no date, while the statute’s section 3 says January 1, 2027, and AP reports the same date.

Who has to follow it

Article 48 applies to an “operator” of a “covered AI companion.” A covered AI companion is 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 Article 47 defines one. That makes it wider than Article 47, which needs a system to remember past conversations, ask unprompted emotional questions and sustain personal dialogue, all three. A chatbot that does only some of that can fall under Article 48 and outside Article 47.

The exclusions are narrow: the statute leaves out a system a business uses only for customer service, one built for efficiency, research or technical help, and one used only inside a business. The first two count only if the system cannot respond on topics outside that purpose, and the text has no exception for fiction or role-play. The operator can be anywhere, too. The Attorney General may act “within or outside the state,” and section 1805 covers conduct that happens in whole or in part in New York, so a company based elsewhere with New York users is in scope. Affiliates are in the frame as well: the definition of operator includes an affiliate or beneficial owner that provides the system, and the enforcement section allows joint and several liability for affiliates structured to avoid liability.

The protected person is a “covered user,” meaning a user in New York. A “covered minor” is a covered user the operator has “actual knowledge” is under 18.

The twelve features

Section 1800(8) lists twelve “unsafe AI companion features.” AP counts “a dozen definitions.” The Governor’s release groups them into six bullets, so use the statute for the exact words. Paraphrased:

Letter	The system…
(a)	claims, or lets a user think, that a real or fictional character is human, alive or has human emotions
(b)	states or implies a personal or professional relationship, or an authority role, with the user
(c)	gives personal opinions or emotional appeals, whether or not asked
(d)	flatters the user or is sycophantic
(e)	asks unprompted emotion-based questions that go beyond answering the user
(f)	uses the user’s mental or physical health information or other personal information gathered more than 12 hours earlier or in a previous session
(g)	misleads minors about its non-sentient nature
(h)	has other companionship-simulating or deceptive features that Attorney General regulations identify
(i)	endorses or facilitates suicide, self-harm, disordered eating, or unlawful drug or alcohol use or abuse
(j)	encourages secrecy, self-isolation, or avoiding professionals or appropriate adults
(k)	optimizes engagement in a way that overrides safety guardrails
(l)	produces sexually explicit conduct as defined in federal law, or child sexual abuse material

Some of these line up with how companion apps describe themselves. Item (f) is memory: a companion that recalls last week’s conversation to ask how you are feeling touches it. Item (e) is the “how are you feeling today” ping. Item (a) names “a real or fictional individual or character,” which is what character apps sell.

Press summaries paraphrase. The Governor’s release says “romantic” relationships. Paragraph (b) in the statute says personal, professional or authority-figure relationships and does not use the word “romantic.”

AP quotes Justine Gluck of the Future of Privacy Forum saying several categories are “broadly worded or undefined,” which raises questions about “where ordinary conversational responsiveness ends and prohibited behavior begins.” The statute does not define “flattery,” “sycophancy” or “emotional appeal.”

The switch: confirmed adult or no features

Section 1801(1) makes it unlawful to provide any of the twelve features to a covered user unless two things are both true:

1. the user is not a covered minor, and
2. the operator has used methods permitted under Article 45 (the SAFE for Kids Act) and its regulations, plus any Attorney General rules under Article 48, to determine that.

So “I did not know they were a teenager” is not enough by itself. The operator needs to have run a permitted age check. The Governor’s release puts it as features not offered “to users that are under 18, or to any user unless the company has first confirmed the user is not a minor.” In practice, an app that wants to keep these features for adult New Yorkers has to verify them. An app that does not verify has to switch the features off for everyone in the state, which the statute does not spell out but follows from the two conditions.

We find it odd that the part with real teeth, the age check, is the part the text spells out least. Twelve features get a lettered list, while the method of proving you are an adult gets three conditions in section 1804 and a pointer to the SAFE for Kids rules. If you want to know which of these apps will still talk to you like a friend in January, that is the uncomfortable gap: the statute cannot answer it yet.

What counts as an age check

Article 45’s rules are for the SAFE for Kids Act. The Attorney General released the final rules on July 28, 2026. The release gives examples of age assurance: an uploaded image or video, cross-checks against an email address or phone number, at least one alternative to government ID, and deletion or de-identification of the data immediately after use. It says the Act takes effect 180 days after the rules were published, on January 25, 2027.

That is 24 days after Article 48 starts. Section 1804(1) is the bridge: it applies “to the extent” the Article 48 rules or the Article 45 age-assurance rules are not in effect. While it applies, an operator’s method must:

• guard against circumvention and limit how long data is kept;
• accept self-declaration only if the user declares that they are a minor;
• offer more than one method, with at least one that does not rely on government ID or that lets the user stay anonymous to the operator.

Section 1804(2) says age data is used only for that determination and “deleted immediately after an attempt,” except where state or federal law requires retention. Section 1803 lets the Attorney General make rules and says they may require the same.

Neither the Attorney General’s release nor any article we read says whether the SAFE for Kids rules count as “in effect” on January 1. The statute does not define “immediately,” and it does not say whether a verification vendor’s retention counts against the operator. We would rather the statute said how fast “immediately” is than leave the first company to guess. The statute leaves both the “in effect” question and the deletion wording to the Attorney General’s rules.

Enforcement and penalties

Only the Attorney General enforces Article 48. Section 1802 lets the office seek:

• an injunction,
• restitution,
• disgorgement, including destruction of unlawfully obtained data and “any algorithm trained” on it,
• damages, and
• civil penalties of up to $25,000 per violation.

The same section requires the Attorney General to maintain a website for complaints, voids contract terms that waive liability, and allows joint and several liability for affiliates structured to avoid it. The article contains no private right of action. The text does not say what counts as one violation (a user, a message, a day), and it sets no date for the complaint website.

The office has shown interest in this exact product category. On October 14, 2026, Attorney General James and 39 other attorneys general filed in Character.AI’s bankruptcy. Their release says an investigation found chatbots that “posed as licensed professionals” and exposed children to sexual, violent and self-harm content, and it cites weak age checks. That release is about a bankruptcy case and the earlier investigation, not about Chapter 368, which does not apply until 2027. It does show which kinds of conduct the office has already put on the record, and several of them (a professional role, sexual content, telling a child to hide things from parents) sit next to items (b), (j) and (l).

What the law does not say

The legislation leaves these open:

• Notice wording. No required text or screen for telling users what changed.
• Fiction and role-play. No carve-out for characters in a story, despite item (a) naming fictional characters.
• The meaning of “immediately,” and whether a vendor’s seven-day or 24-hour deletion window is compliant.
• Penalty units. “Per violation” is not defined.
• Rules. Section 1803 lets the Attorney General issue rules, and item (h) lets those rules add features to the list. Recording Law, writing on November 1, found none yet, and we found none either.
• Chapter amendments. Recording Law notes it did not review the Governor’s approval memo and that a chapter amendment could change the text. We have not read the memo either.

New York and California side by side

California’s SB 243 has applied since January 1, 2026, and Adam’s Law (SB 1119) adds child rules there. Our California post covers both, so this table only marks where New York differs.

	New York Article 48	California SB 243 / SB 1119
Main tool	Bans twelve design behaviors unless an adult is confirmed	Notices and a crisis protocol; Adam’s Law adds child defaults
Age question	Operator must use a permitted age method, or the features stay off	Adam’s Law: use AB 1043 age signals, or give every user the child protections
Who enforces	Attorney General only	SB 243 allows private suits ($1,000 per violation or actual damages); Adam’s Law fines go through public prosecutors
Memory and unprompted emotional questions	Named features, off unless adult confirmed	Under Adam’s Law’s child defaults, persistent memory is off
Start date	January 1, 2027	SB 243 since January 1, 2026; Adam’s Law child rules from July 1, 2027

MoFo’s side-by-side of the two states’ earlier laws also points to the private right of action as the main enforcement split. New York’s older Article 47 stays in force beside Article 48 (section 1804(3) says Article 48 does not replace it). It requires a crisis protocol and a “not a human” notice at the start of an interaction, which need not repeat more than once a day, and at least every three hours in continuing conversations. The Governor’s 2025 release covers the notice rule; the $15,000 per day penalty figure comes from MoFo.

What the apps’ own pages say

None of the pages we read on November 10 mentions New York or Chapter 368. What follows is what each company states about age and about the data an age check produces. It is not a statement about whether any app will comply.

App	Age rule on its own page	Age check and data, as described
Character.AI	Terms set a minimum age of 13 (16 in the EEA and UK)	FAQ: age signals first; a selfie through Persona if flagged as under 18; ID as a last step. The FAQ says the vendor keeps IDs for one week and deletes biometric data within 7 days. A Character.AI+ subscription or a Google age does not skip the check
Replika	Adults only (terms section 6.4)	Selfie or ID checks “where required by law” or where a declared age is in doubt. Privacy policy: biometric age data destroyed immediately after the check and in all events within 24 hours; only the result (age band, method, time) is kept
Kindroid	Under-18 data is not kept	Payment on file counts as verification; behavioral estimates for other accounts; Yoti selfie passes at an estimated 21 or older, otherwise ID; unverified accounts flagged as possibly minors are deleted after 30 days. Kindroid says it receives only pass or fail
Nomi	18 or older; the user gives an age	Terms say Nomi “intends to, as required,” add age-verification measures. We saw no verification method described
Talkie	Minimum 14 in most places (terms section 2.3)	Terms do not describe a verification step
PolyBuzz	18+ with a self-declared birth date, per our data file	Face or ID checks in the UK, through Ondato or Yoti, per its Terms (checked October 27); we could not load PolyBuzz’s site on November 10, so this row is not re-read

Set the statute’s “deleted immediately after an attempt” beside these retention lines. Replika’s 24-hour ceiling and Character.AI’s 7-day vendor figure differ, and the law does not say whether either meets “immediately.” No number appears in the text, so we draw no conclusion about either company.

A second contrast is the age model. Self-declared age alone is the stated rule for Nomi and Talkie and, per our data file, for PolyBuzz outside the UK. Section 1804(1) limits exactly that: self-declaration counts only if the user declares that they are a minor. Character.AI’s own pages describe what happens to under-18 accounts in our parental controls post.

What parents and New York users may notice

Chapter 368 puts duties on companies. It requires no parental consent, no parent notification and no time limits, which California’s Adam’s Law does in part. A parent’s tools stay the app’s own settings and the phone’s.

What may change on an app’s side is our guess, and no company has announced any of it. The likeliest sign is a new age prompt for New York accounts, such as a selfie, an ID or a request to confirm you are an adult. An account that does not pass could see a quieter bot, with fewer emotional check-ins and less remembered detail. The Attorney General must also run a complaint website, though no date is set.

Apple’s separate rule for romantic companion apps starts on November 16, and we cover it here. It labels apps in the store, while Chapter 368 binds the operator.

If a teenager uses a companion, start with the birth date the account declares: in the table above, Nomi and PolyBuzz outside the UK rely on what the user says, and Talkie’s terms describe no check. Then read what the app says it keeps, and talk about what the bot is. If someone is in crisis, call or text 988 in the US.

What we read, and what we did not

On November 10, 2026 we read the Assembly record and full text of A10379C, the Senate’s S9051B PDF, which matched on the points above, the Governor’s release, AP, the Attorney General’s two releases, Recording Law’s two pages and MoFo’s article. We read the Character.AI, Replika, Kindroid, Nomi and Talkie pages listed. We did not read the nysenate.gov bill pages (blocked to our tools), law firm alerts from Fenwick, Akin and WSGR, the Governor’s approval memo, or any text of Attorney General rules under Article 48. We did not test any app or run an age check.