Every companion law written in 2026 draws its line at eighteen. Common Sense Media’s new census finds that 58 percent of American children aged nine to twelve have already used an AI chatbot, that one in six chatbot users has been shown something not meant for their age, and that most of them did not tell an adult. The age the law protects is not the age that is talking.
The Common Sense Media Census: AI Use by Tweens and Teens, 2026, surveyed 1,204 American children aged nine to seventeen in March 2026, recruited through their parents on a national probability panel and weighted to population benchmarks. The headline number is that 67 percent have used an AI chatbot. The number underneath it is the one that matters: 58 percent of nine-to-twelve-year-olds have, rising to 71 percent of thirteen-to-fifteen-year-olds and 77 percent of sixteen-to-seventeen-year-olds. Daily use of any AI runs from 17 percent among the youngest to 30 percent among the oldest. The most-used product, by a wide margin, is ChatGPT, named by 40 percent; Google’s offerings come second at 24 percent.
Then the safety figures. Seventeen percent of chatbot users say a chatbot has said or shown them something that was not OK or not meant for their age; 8 percent say it has happened more than once. Among nine-to-twelve-year-olds the figure is 16 percent. Of the children this happened to, a third told a trusted adult about the most recent incident. More than half did not.
And 44 percent of all children surveyed say no parent or guardian has ever talked with them about using AI safely. Among daily users, a third have never had that conversation.
The age nobody legislated for
Read those numbers against the regulatory year. California’s SB 243, New York’s two companion laws, the federal GUARD Act, Apple’s new 18+ rating, the EU’s Kids Act: every one of them is built around the line between a minor and an adult, and in practice around teenagers, because teenagers are who the lawsuits were about. The design assumption is a sixteen-year-old with a phone and a romantic chatbot.
The census describes a different population. A ten-year-old does not have a companion app. A ten-year-old has ChatGPT, or Gemini on the family tablet, or whatever the school handed out, and uses it for homework, and then for the other things a ten-year-old wonders about. Fifty-eight percent of them are already there. The laws that require age assurance before a companion feature switches on assume the child is trying to get into a product designed for adults. The census suggests the child is already inside a product designed for everyone, and the companion features, the memory, the warmth, the follow-up question, are simply what the product does.
New York’s new law, to its credit, reaches general assistants, and its features list would switch most of this off for a verified minor. But it takes effect in 2027, in one state, and it depends on an operator knowing the user is a child. The census found that a quarter of these children use AI daily without their parents ever having raised the subject. The operator is not going to know what the parent does not.
What “not meant for my age” means at eleven
The 17 percent figure deserves a pause, because it is self-reported by children, and children are conservative reporters of what they were not supposed to see. One in six chatbot users between nine and seventeen says the chatbot itself, not another user, not a search result, produced something inappropriate for them. For the nine-to-twelve group it is one in six as well.
The companion industry’s defense has always been that its products are rated 17+ and that children lie about their age to get in. The census covers no companion apps at all. It covers the chatbots children are allowed to use, the ones with 12+ or 4+ ratings and school district contracts, and it finds that those produce inappropriate content for a sixth of their child users. The problem the laws are chasing in companion apps turns out to be a property of the conversational interface itself, at every age and in every product, because a system that generates text in response to a child will eventually generate text a child should not see. Apple can rate the romantic ones 18+. It cannot rate ChatGPT.
The 53 percent who said nothing
The number with the longest consequences is the smallest: of the children shown something not meant for them, 53 percent did not tell a trusted adult, and 14 percent preferred not to say whether they had. Combine that with the 44 percent who have never discussed AI safety with a parent, and a pattern that runs through all of 2026’s youth research sharpens. Two thirds of young people who ask an AI about their mental health tell no one. More than half of children shown something inappropriate by a chatbot tell no one. The conversation with the machine is private in a way the conversation with the parent is not, and children have noticed.
That is a design outcome rather than an accident of childhood. A chatbot never reacts the way a parent might, never looks worried, never asks a follow-up that makes the child regret saying it. For a child those are features. They are also exactly the properties that keep the adult out of the loop, and the census shows the loop closing around nine-year-olds.
What a counterpart for children would be
There is an argument, which Common Sense itself makes, that none of this means children should not use AI, and the census supports it: 40 percent of child AI users practice conversations and social skills with it, and among children who find making friends hard, nearly half have talked to it about their feelings. Something is being met there. The question is what meets it, and whether it hands the child back to the people who should.
A counterpart built for adults, from a real adult, is not for a ten-year-old and should not pretend to be. Someone to think with presupposes a person who has already learned how to think with people, and a child is still learning that from the people around them. What a child needs from a machine is the opposite of what the census shows children getting: something that reacts, that has an adult’s judgment about what a ten-year-old should see, and that treats a hard disclosure as a reason to say “who else knows about this?” rather than as a conversation to continue. Growing up needs friction, and the one friction a child cannot do without is an adult who finds out.
The laws will keep drawing the line at eighteen, because that is where law can draw it. The census has moved the problem to nine. Fifty-eight percent of them are already talking, and most of their parents have not asked about what.
Sources: Common Sense Media, The Common Sense Media Census: AI Use by Tweens and Teens, 2026 (fieldwork March 18 to 26, 2026, SSRS Opinion Panel, N=1,204 children aged 9 to 17, margin of error plus or minus 3.8 points; chatbot use by age, daily use, product shares, inappropriate-content figures, parent-conversation figures, social-skills and feelings figures). Common Sense Media, The 2026 Common Sense Media Census: Tech and Media Use by Tweens and Teens (October 27, 2026). NPR via Benton Institute (October 30, 2026).
