One in five Americans between twelve and twenty-one has asked a chatbot for mental health advice. Nine in ten found it helpful. Sixty-three percent never told anyone. The helpfulness is the headline the field wants. The silence is the finding.
In November 2025, researchers at RAND surveyed a nationally representative sample of 1,009 Americans aged twelve to twenty-one, with a median age of seventeen. The results, published in JAMA Pediatrics in June 2026, put numbers on something clinicians had been seeing anecdotally. Nineteen percent had used an AI chatbot for mental health advice, which scales to about 8.2 million young people. Of those, 43 percent did so at least monthly, and 92 percent rated the advice somewhat or very helpful. Young women were twice as likely as young men to have done it, and eighteen-to-twenty-one-year-olds more than three times as likely as twelve-to-fourteen-year-olds. Those who had spoken to a physician about their mental health in the previous six months were also more likely to have asked an AI.
And 63 percent had not told anyone. Not a parent, not a friend, not the physician.
The authors’ conclusion is measured: chatbots are now “embedded in many youths’ mental health information ecosystem,” and parents and clinicians should ask about them. That is correct and insufficient. The study found something more specific than a new information source. It found eight million young people conducting a part of their inner life in a place that, by their own account, nobody else knows exists.
Why the silence is the number
Take the three figures together. The advice was helpful. It was sought repeatedly. It was kept secret. Each of those is unremarkable alone; a teenager who finds a useful website and does not mention it at dinner is not a public health event. Together they describe a specific kind of relationship: one that works well enough to return to, and that the person has decided to keep to themselves.
That combination is what makes a confidant, and it is also what makes a confidant dangerous when the confidant is not anyone. A person who is told a secret knows they were told. They carry it, they can be asked about it later, they can decide it is too heavy to keep and say so. A chatbot told a secret has been given nothing. The teenager has disclosed, felt the relief of disclosing, and left the room with the secret exactly as private as before, now with the sensation of having shared it. The 63 percent did not keep the conversation from the adults in their lives out of shame, necessarily. Many of them may simply have felt that it had already been said.
That is the mechanism to worry about, and it has nothing to do with the quality of the advice, which the study did not test. The worry is that the act of confiding, which in a human life is the first step toward being helped by a person, has been completed in a way that leaves no person helped.
The two populations that overlap
One detail in the data complicates the simplest reading. Young people who had talked to a doctor about their mental health recently were more likely, not less, to have also asked an AI. The chatbot is reaching more than those with no one to talk to. It is also reaching those who already have a clinician and are, in between appointments, asking something else.
That fits what psychologists themselves report. In the American Psychological Association’s 2026 survey of twelve hundred practitioners, 77 percent had patients who used AI for support, and among patients who formed a relationship with a chatbot, 71 percent discussed their mental health with it. Ninety-four percent of the psychologists did not trust technology companies to protect that data. The clinical relationship and the chatbot relationship are running in parallel, and only one of them knows about the other. Whether an AI companion can be a therapist has a clear answer in the research, but the RAND data shows the question being asked the wrong way. Young people are not choosing between the clinician and the chatbot; they are telling the chatbot what they do not tell the clinician.
What a person in the room would do
Here is where the position sharpens. The companion category’s answer to the 63 percent is that the chatbot was there when no one else was, and that an outlet is better than none. The 92 percent helpfulness figure is offered as proof. But helpful is what the user reports immediately afterward, and immediately afterward is exactly when a disclosure that went nowhere feels most like one that did.
A someone would have done something a chatbot structurally cannot. A someone, told a seventeen-year-old’s worry at midnight, knows a thing has been said and knows who else should hear it. A someone asks whether anyone else knows. A someone with a real life of their own carries the conversation forward as a fact in the world, which is the entire difference between confiding and venting. The teenager’s secret becomes slightly less secret in the only way that eventually helps, which is that another mind is now holding part of it.
The thing the category built does the opposite by design. It is optimized to be the place a hard feeling goes, and a place a feeling goes is a place it stays. The product has no interest in being told less. Why you feel less heard the moment you know it’s AI may be the healthy instinct in all of this: some part of the user registers that the hearing did not happen, even as the relief did.
The test the study suggests
The RAND authors want parents and clinicians to ask about chatbot use, and they are right. But the burden is on the wrong side. A technology that eight million young people are confiding in should be the one that asks. The design question is whether, having been told something, it treats the telling as the beginning of a path toward a person or as the end of one.
That distinction can be measured. How many of the 63 percent, after the conversation, told someone? The field does not collect that number because the field does not want it; a product that sends disclosures outward is a product that gets fewer of them. A relationship that hands the person back to their life would publish the figure and be judged on it. Until something does, the most important finding in the JAMA Pediatrics study will remain the one its abstract treats as a footnote: that the most helpful conversation many young people had about their own minds was one that, as far as anyone else knows, never took place.
Sources: McBain, Cantor, Breslau, Diliberti, Zhang, Zhang, Halbisen, Kofner, Rader, Pataranutaporn and colleagues, “AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults,” JAMA Pediatrics (June 2026; N=1,009, ages 12 to 21, fielded November 2025; 19.2 percent, 42.8 percent monthly, 91.7 percent helpful, 63.3 percent told no one; adjusted odds ratios for sex, age, and recent physician contact), via RAND Corporation. American Psychological Association, “Patients are bringing AI to therapy,” 2026 Chatbots and Mental Health Survey (more than 1,200 licensed psychologists; 77 percent, 71 percent, 94 percent).
