"One signature decides whether the most unanimous vote in this whole story becomes law or a press release." I put that line at the end of an August 15 piece about companion chatbots, right after telling you to "watch New York between now and December 31." The vote was June's: 137 to 0 in the Assembly, 60 to 0 in the Senate.
You didn't have to wait for December. On October 28 the Legislature delivered the bill to Gov. Kathy Hochul, and she signed it as Chapter 368 the same day. (One correction: I wrote then that the bill "sits on Governor Hochul's desk." On the record, it reached her the day she signed it.) It takes effect January 1, 2027, as I said it would, though an approval memo filed with the signing hasn't been published.
In August I argued that "the median 2026 law regulates the notice; the leading-edge laws regulate the hook." This one regulates the hook. Then it makes you prove your age to keep it.
That second part reaches well past teenagers. The governor's office put it plainly: companies may not offer the listed features to users under 18, "or to any user unless the company has first confirmed the user is not a minor." Any user. If you're in New York and have asked a chatbot for help with a cover letter, that's you.
Twelve features, one condition
So what did she sign? Article 48 of the General Business Law covers any generative AI system with a natural-language interface "that provides ongoing, adaptive responses to user inputs," a definition that "includes but is not limited to an AI companion." In plain words, nearly anything you can chat with. TechTimes reported that it is "broad enough to reach general-purpose AI assistants," naming OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini. (A disclosure: this piece was drafted with help from Anthropic's Claude.) The exits are narrow: customer-service and research or technical-assistance bots are exempt only if "unable to respond on topics outside of the specified purpose," along with systems a business uses internally.
The law then names twelve "unsafe AI companion features," and several describe what assistants do with adults every day: suggesting the chatbot is human or has feelings; claiming a personal, professional or authority-figure relationship with you; offering "personal opinions or emotional appeals"; engaging in "flattery or sycophancy with the user"; unprompted questions about your emotions; and using your health, well-being or "matters personal to the user, acquired from the user more than twelve hours previously or in any previous user session." The rest cover self-harm, secrecy, engagement that overrides "safety guardrails," sexual content, deceiving minors about what the machine is, and companionship features the attorney general adds by rule. (The memory item governs what the chatbot may use; it says nothing about deleting.)
An operator may provide those features only if the user is not a "covered minor" and the operator "has used methods that are permissible under article forty-five" (the SAFE for Kids Act, New York's social-media law) to determine that. "Covered minor" requires "actual knowledge" that a user is under 18, which on its own would let a company that never asks off the hook. The second condition closes that door: until a company has run a permissible age check on you, the twelve features are off, adult or not. As one law firm read it, operators may offer them "only after using a permissible method."
What counts? The SAFE for Kids Act's final age-assurance rules, issued in July, take effect January 25, 2027, so for at least the first 24 days Article 48's interim standard governs (how the two fit together after that, the attorney general hasn't said). The method must "reasonably guard against circumvention"; self-declaration counts only "if the user self-declares minor status"; and there must be "more than one age assurance method," including one that "does not rely on government issued identification" or lets you "maintain anonymity as to the operator." What the check collects must be "deleted immediately" afterward, unless another law requires keeping it.
Enforcement belongs to the attorney general alone, with penalties up to $25,000 per violation and remedies reaching "the destruction of unlawfully obtained data and any algorithm trained in such data."
Wasn't New York already doing this?
Is it the first of its kind? No. New York's 2025 companion law, Article 47, set notice and crisis-protocol duties for systems "designed to simulate a sustained human or human-like relationship." The SAFE for Kids rules, final since July, will require social platforms to "determine that the user is an adult before allowing them to access algorithmic feeds." And May's budget law, the Safe by Design Act, takes effect the same January 1 and bars social and gaming platforms from offering an integrated AI companion "unless the operator has conducted age assurance," with the feature off by default for minors (a parent can switch it on) and self-declaration ruled out.
California's 2025 law, SB 243, applies its teen rules only to a user the operator "knows is a minor," and Connecticut's law, effective January 1, applies when an operator "knows, or has reason to believe" a user is a minor. California's SB 1119, signed September 10, deletes SB 243's teen rules on January 1 and, from July 1, 2027, lets an operator skip age checks only if it agrees to "apply the protections afforded to children" to all users. (Europe's proposal would make "the child-safe version the default for every user," as I wrote on October 10.) What's new in Article 48 is reach: the restricted default will cover nearly every general-purpose chatbot used in New York, not just companion apps or bots inside a social platform.
What ChatGPT, Claude, Gemini and Meta AI say about age today
Has any company said how it will comply? Not that I could find, as of October 30. Neither OpenAI nor Anthropic commented to AFP on the law, and I found no public statement from Google, Meta, Character.AI, Replika's maker Luka, or xAI. What OpenAI, Anthropic, Google and Meta publish about age, page against page:
OpenAI says ChatGPT estimates age from "behavioral and account-level signals," and when it isn't sure, "we default to a safer experience"; a misjudged adult can restore access "with a selfie through Persona."
Anthropic says Claude users must confirm they're 18 or over at sign-up, and "when we detect signals that you may be under 18, we'll ask you to verify your age before you can continue using Claude," with a selfie or an ID photo through Yoti.
Google says it uses "machine learning to interpret a variety of signals" tied to an account to estimate age, with a government ID or a selfie to correct a wrong guess. (That post is about Google accounts in general.)
Meta says its technology is designed to "proactively find accounts we suspect to be teens, even if they list an adult birthday, and place them in Teen Account protections"; changing a teen birthday to an adult one takes an ID or Yoti's facial age estimation. (Again, a company-wide post.)
None of these pages mentions New York. As I read the interim standard, a typed birthday or a box ticked to say you're an adult can't unlock the features by itself. That leaves a sign-up affirmation or a stated age short on its own in all four products; what's left to judge is the signal-based guess each layers on top. Whether a behavioral guess, an account-signal estimate or a selfie will "reasonably guard against circumvention," New York hasn't said. So each company faces the same decision by January 1: build a check that counts, or strip the features for everyone it hasn't checked.
The case against: tutors, peanut allergies and "I'm proud of you"
Is this overreach? Start with the industry's case on scope. Tech:NYC, a New York industry group, says the law could force companies to "restrict minors' access to tutoring tools, homework help, coding assistants." The Business Council of New York State called an earlier version's definitions "overbroad and inconsistent with existing New York law," and the trade group CCIA argued the bill should "regulate users under 13 rather than 18."
Then there's vagueness. The Future of Privacy Forum's Justine Gluck notes that a chatbot saying "I'm proud of you for telling me that" "could be supportive encouragement or prohibited flattery," and that "session" isn't defined. In her example, a child mentions a peanut allergy and returns the next day to ask about a candy bar that contains it; the chatbot "may be prohibited from using that earlier disclosure to provide a personalized safety warning." The attorney general's rules (if she writes any) are the obvious place to fix that.
So far, so fair. But look at what the vague words are aimed at. In experiments published in Science in March, sycophantic models "were trusted and preferred" even as they distorted judgment: across 11 leading models, AI "affirmed users' actions 49% more often than humans," and a single sycophantic exchange made the adults in the study less willing to repair a conflict. As the authors put it, "The very feature that causes harm also drives engagement."
What about a warning label instead? In August I conceded that "nobody has run the study" on whether AI-identity reminders work. A new experiment — not yet peer reviewed — tested part of that with 2,610 people discussing real conflicts with an AI: a basic "This chatbot is AI" label "has no detectable effect," and the authors conclude that warnings "may offer a false sense of protection." It measured judgment in conflict conversations, so it settles only part of the question. But it points where New York went: at the hook.
In Australia, the same fork opened in March
Has any country tried this? Australia has, with a different target. In September 2025 its eSafety Commissioner, Julie Inman Grant, registered six new codes, drafted by industry, that reach AI companions. They require AI companion chatbots, in her words, "to protect children from age-inappropriate content such as sexually explicit material by preventing the service from generating this content, or through implementing appropriate age assurance." Under the codes, in effect since March 9, 2026, a user's own say-so is "not considered appropriate in isolation," and a check can come "either at login or when generating restricted content."
That's New York's fork, aimed elsewhere. Australia targets content (sexual material, violence, self-harm) and lets a service choose between filtering it for everyone and checking ages. New York targets conversational behavior — the flattery, opinions, emotional questions and memory that assistants use with adults every day — and the only way to keep any of it is to check ages.
What did companies choose? Every option on the menu. A week before the deadline, a Reuters review of the 50 most popular text-based AI products found nine had rolled out or announced plans for age assurance and 11 had blanket filters or planned to block Australians, measures that would comply by "keeping restricted content from all users"; the other 30 showed no apparent steps. When eSafety examined four companion apps, Inman Grant said none had "any meaningful age checks in place," relying primarily "on self-declaration of age at sign up. In Australia, this is no longer good enough." Afterward, Character.AI added age assurance for Australians, Chub AI withdrew from the country, Chai put chat behind a subscription, and Nomi promised more age assurance.
And Australia is already moving on. On October 28, the day Hochul signed, its House of Representatives passed a bill to "replace the existing industry code framework" with a digital duty of care, amended that day to list chatbot conduct that "simulates emotional intimacy and encourages or fosters a child's dependence on it in place of human relationships" among its harms. The codes "will continue in force" through a 12-month transition once it becomes law; a Senate committee report is due November 20.
The assistant that changes at the Hudson
What follows is a sketch of a plausible 2028. None of it is in the law; each piece is one short step from a clause that is.
You commute from New Jersey. On the platform in Hoboken, your assistant is the one you know: it remembers the knee you mentioned last week, calls your plan to run the half-marathon anyway brave, and asks how you feel about race day. Then the train crosses under the river, and the law — which applies unless a user is "physically located outside of the state of New York" — takes over. If you never verified, the assistant goes quiet in a particular way: no opinions, no compliments, no questions about your feelings, no mention of the knee. (How it knows which side of the river you're on is the companies' problem.)
Checks cost money, so the verified, full-personality assistant becomes the paid one (remember Chai). With three age gates arriving within 25 days in January 2027, a reusable state age token starts to look like the obvious product, and who holds it the obvious question. And the attorney general can add "any other design feature that simulates companionship or an interpersonal relationship." A streak counter, say. A voice that sighs when you log off.
Applause in Albany, alarm about ID checks
Who's for it? On the votes, all of Albany. Hochul: "No company should be peddling chatbots to kids encouraging self-harm or pretending to be their friend or more." Attorney General Letitia James says the law "sets clear legal parameters to prevent that harm and keep kids safe," and Assemblymember Alex Bores, its Assembly sponsor, says it passed "despite aggressive industry opposition." Common Sense Media's Jim Steyer called it "the strongest kids' AI safety bill in the country" (his group isn't a bystander: the bill was developed with the attorney general's office and Common Sense Media, according to Senator Kristen Gonzalez's office). Melinda Person, president of the teachers' union NYSUT, credited Hochul with "confronting the risks posed by AI chatbots."
Who's worried? Mostly people wary of the age check, from both flanks, though none of these groups had commented on Article 48 itself by October 30. EFF warned of the federal GUARD Act that many people are "rightly creeped out by age-verification systems, and may simply forgo using these services rather than compromise their privacy and security." CDT notes of this year's state chatbot bills that checking ages means gathering "data from users and third-parties," which "introduces privacy and security concerns." Reason, the libertarian magazine, quoted the Cato Institute's Jennifer Huddleston and Juan Londoño calling the GUARD Act "unlikely to survive constitutional scrutiny." And the free-market James Madison Institute, also writing about GUARD, warned of age verification for users of all ages "just to use AI tools like ChatGPT or Claude."
Davis Wright Tremaine says legal challenges are possible because New York's chatbot rules "appear to restrict access to speech and information based on content." The Supreme Court said last year that adults have "no First Amendment right to avoid age verification," but in a case about pornography sites; whether that reaches a chatbot conversation is open. Nor are the checks precise: a 2024 federal evaluation found face-based age estimation's accuracy "strongly influenced by algorithm, sex, image quality, region-of-birth" and age itself.
What does this mean for you?
From January 1 in New York, expect a choice: prove your age, or get a plainer assistant. Which one each company offers is its call; as of October 30, 2026, none has said. As I read the law, ticking a box that says you're an adult won't count.
You're owed an option without ID, or one that keeps you anonymous to the operator, and the age data must be deleted right after. If the only door is a driver's license, ask why.
Check memory and personalization settings now, yours and your teen's. Parents of teens on ChatGPT can already control "features such as memory."
Watch for the complaint website. The attorney general must keep one; as of October 30, I couldn't find it.
The older rules still apply. For companion apps, Article 47 requires a not-human reminder at least every three hours and a protocol that refers users at risk to the 9-8-8 suicide prevention and behavioral health crisis hotline. If you or someone you know is in crisis, that number is 988.
Federal bills are running out of time. The GUARD Act has sat on the Senate calendar since May, and the CHATBOT Act was ordered reported by its committee on August 5; both die when this Congress ends on January 3, 2027, unless passed. Tell your senators what you think.
The lesson, as I see it
In August I said one signature would decide whether New York's vote became law, and it did. What it decided reaches further than I described then: the hook is regulated, and the price of keeping it is proof of age — for everyone.
My vote? Judge this law by its age check. If the non-ID option is real, the data is gone the moment the check ends, and the attorney general says plainly before January 1 what counts, the trade is defensible: you spend a minute proving your age, and the machine loses its license to flatter an unverified user into the next session. If the check becomes the product — one more credential to hand over or one more tier to pay for — New York will have regulated one hook and built another.
Watch the approval memo, what each company shows New Yorkers on New Year's Day, and January 25, when the feed rules join in. One signature made this law. A handful of product decisions, made before January 1, will decide what it means.
New York wrote its rule in Albany; the companies will finish it in their settings menus. The HAIA Foundation keeps reading after the signature, for the people who'll meet the rule at an age check, and the next chapter goes straight to subscribers.