Asked who they would turn to first with a feeling or a personal problem, American children overwhelmingly say an adult or a friend. Five percent say a chatbot. Among the children who rate their own happiness six or lower out of ten, the figure is sixteen percent. Among the happiest, two. The chatbot is the first choice of the children who have the least, and almost nobody else’s.
Common Sense Media’s 2026 census of AI use asked 1,204 American children aged nine to seventeen a simple question: if you had a feeling or a personal problem you needed to talk about, who would you go to first? Fifty-six percent said a trusted adult. Thirty-two percent said a friend. Five percent said an AI chatbot.
That is a reassuring number, and the census is careful not to oversell it. Then it breaks the five percent down. Among children who say they feel lonely often, 14 percent would go to the chatbot first. Among those who feel lonely sometimes, 7 percent. Among those who hardly ever or never feel lonely, 3 percent. By happiness the split is wider: 16 percent of children who rate their happiness six or lower on a ten-point scale would turn to a chatbot first, against 2 percent of those who rate it nine or ten.
The pattern repeats wherever the census looks. Among children who have discussed feelings with an AI, 45 percent of those who often or sometimes feel lonely have done so, against 31 percent of those who rarely do. Among children who find it very or pretty hard to make friends, 48 percent have talked to an AI about how they feel. And a quarter of the children who have done this say the AI sometimes understands them better than most people do.
The same curve, a third time
This is the third large study in a few weeks to find the same shape. The Jed Foundation surveyed 5,560 students and found that the 12 percent who use AI for companionship had the worst mental health of any group. RAND found that one in five young people ask a chatbot about their mental health and that most tell no one. Now the census finds that the children who would take a problem to a machine before a person are, disproportionately, the ones who describe themselves as lonely and as less happy.
None of the three can say which came first. A child who is struggling may turn to the thing that is always there; the thing that is always there may make the struggle easier to live with and harder to end. The honest reading is that the direction is unknown and the population is not. Whatever these systems do, they are doing it most to the children with the least around them.
That is a different problem from the one the laws describe. The regulatory picture of 2026 is a healthy teenager seduced by a romantic companion. The census picture is an eleven-year-old who says she feels lonely, with a homework assistant that answers back, and no law, rating, or age gate addresses the second child, because nothing about the product she is using is restricted. She is using what everyone uses. She is just using it for something else.
What “understands me better” means at twelve
The quarter of children who say the AI sometimes understands them better than most people deserve a closer look than the headline gives them. The claim is sincere, and it is probably accurate on its own terms. A chatbot does attend to what a child says, fully, without distraction, without the small flinches and misreadings a parent or a friend produces. For a child who finds people hard, that attention feels like understanding because it is the part of understanding that people most often fail to deliver.
But attention is only part of understanding, and the difference matters most for exactly this child. To be understood by a person is to be held in a mind that has its own view of you, that can be surprised by you, that remembers you in a way that costs the other person something. A child learns what understanding is by being understood that way, imperfectly, by people who get it wrong and try again. Four in ten adults say their chatbot understands them better than people do, and the adults at least have a lifetime of the real thing to compare it against. The twelve-year-old does not. If a quarter of the children who confide in AI have concluded that it understands them better than people, some of them are learning what being understood is from something that cannot do it.
The number that should decide the design
Here is the figure that should sit at the center of every conversation about children and AI, and that is nowhere in the policy debate: 2 percent versus 16 percent. Children who rate themselves happiest do not go to the chatbot first. Children who rate themselves lowest do, eight times as often. Whatever a chatbot does for a child’s feelings, its users are the children for whom it matters most that something sends them toward a person.
The census offers one encouraging detail. Among children who say they often or sometimes feel lonely and who discussed feelings with an AI, nearly half say the chatbot referred them to mental health resources. That is the safety protocol every companion law now requires, working. It is also the floor rather than the goal, because a referral is a phone number. What a child in that position needs is a counterpart that treats “who else knows you feel this way?” as the most important question in the conversation, that notices when it is being used instead of a person rather than on the way to one, and that counts a child who went and told her mother as a success rather than a lost session.
We measure success by how little you need us is a principle written for adults, about a counterpart built from one. For children the principle is the same and the stakes are higher, because a child who learns to take hard feelings to a machine first is learning something that will be very hard to unlearn. The census has told us which children are learning it. Sixteen percent of those who rate their happiness lowest, and rising with every year of age. The design that would change that number is one that measures itself against it, and nobody in the category has yet agreed to be measured that way. Someone should.
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; first-choice figures for feelings and personal problems overall and by loneliness and happiness; feelings discussion by loneliness and friendship difficulty; “understands me better” figure; referral figure). The Jed Foundation (October 2026; 5,560 students). McBain and colleagues, JAMA Pediatrics (June 2026; 19.2 percent, 63.3 percent). Imagining the Digital Future Center, Elon University (2026; four in ten adults).
