The largest school-based study of teenagers and AI yet published sorted students by what they use it for. The twelve percent who use it for companionship had the worst mental health of any group, on every measure, including a history of suicidal thoughts and attempts. The researchers are careful to say they do not know which way the arrow points. The number is worth sitting with either way.
The Jed Foundation surveyed 5,560 middle and high school students at 21 American schools between October 2025 and April 2026, with input from the Digital Wellness Lab at Boston Children’s Hospital. The full report arrived in late October 2026. Four in five students had used AI. Fifty-eight percent used it for homework. About 28 percent used it for advice about relationships or problems. Fifty-seven percent had turned to it when sad, stressed, lonely, or overwhelmed. Twelve percent used it for friendship or companionship.
Then the researchers lined the groups up against measures of anxiety, depression, loneliness, and lifetime history of suicidal thoughts and attempts. Every use beyond homework tracked with worse outcomes than not using AI at all, and homework users themselves reported more lifetime loneliness and suicidal thoughts than non-users. The companionship group sat at the bottom of every measure. Middle schoolers, girls, and students who are American Indian or Alaska Native, Black, Hispanic, or multiracial were significantly more likely to turn to AI when struggling.
Kayla Anderson, who led the research, said what the data permits: AI “may offer an immediate and accessible outlet for young people in distress,” and “it cannot replicate the developmental, protective, and necessary benefits of genuine human connection.”
What the study does and does not show
Start with the limits, because the finding is strong enough to survive them and weak enough to be misused without them. The sample is large but not national; the schools were already enrolled in JED’s programs. The design is cross-sectional: one survey, one moment. The authors say plainly that “it is unclear whether AI use contributes to poorer well-being or whether students already in distress are more likely to turn to AI.” A teenager who is already lonely is more likely to open a chatbot at midnight, and the survey cannot tell that teenager from one the chatbot made lonelier.
So the honest reading stops short of saying AI companionship harms teenagers. What it can say is that the teenagers who are doing worst are the ones using AI as a friend, and that the companionship category has, whether it caused it or not, found exactly that population. Twelve percent of students, and the twelve percent most in need of a person.
The same twelve percent, twice
The number has a twin. Earlier in October 2026, a nationally representative survey of American adults from the University of Iowa found that 12 percent of AI users turn to it for companionship or conversation. Among those adults, 58 percent believe it solves their problems better than a therapist does. The two studies measured different populations with different instruments and the coincidence of the figure should not be overread. But set them side by side and a shape appears. Roughly one in eight people who use these tools, at any age, has moved them into the role of company. Among adults, that group rates the machine highly. Among teenagers, that group is doing worst.
The difference may be the people. It may also be what a companion does to someone who has not yet built the life it is substituting for. An adult who talks to an AI at night has, usually, a day full of people the AI is competing with. A fourteen-year-old who does the same may be doing it instead of learning how to have the day at all. Growing up needs friction, and a counterpart that never provides any is a different thing in the hands of someone who still needs to practice.
Fifty-seven percent
The companionship figure gets the headline, but the larger number may matter more. Fifty-seven percent of students had turned to AI when they were sad, stressed, lonely, or overwhelmed. That is not the twelve percent who call it a friend. That is most of them, using a tool built for homework as the first place a hard feeling goes.
This is the pattern the whole youth research keeps finding from different angles. One in five American adolescents and young adults uses a chatbot for mental health advice, and nearly two-thirds of them have told no one. Swedish children, in a national helpline’s reporting, talk to an AI at night about what they cannot say in daylight. For most of these students the machine takes something smaller than a friend’s place. It takes the moment before a friend, the one where you decide whether to tell someone, and it is winning that moment because it is always there and never has a reaction.
That moment is where a relationship with a person begins. A teenager who takes every hard feeling to something with no reaction is practicing a version of disclosure that never costs anything and never builds anything. The feeling gets processed; the person who could have received it never finds out it existed.
What the finding asks of the people building these things
A position worth stating plainly. The twelve percent who use AI for companionship are the companion industry’s indictment rather than its success, and the JED data is the clearest version yet of why. Whatever the direction of causation, these are the students most in need of a human being, and the product that found them is one that will never hand them to one. A companion built to be returned to has no reason to say, to a fourteen-year-old at midnight, that this is a conversation for a person who can see her face.
The design that would change the number is not hard to describe. It notices the fifty-seven percent moment and treats it as the point. It asks who else knows. It measures itself by whether the student talked to someone with a pulse the next day, and it counts that as a win rather than a leak. Measuring success by how little someone needs you is a strange thing for a technology company to do, and the JED numbers are the argument for saying it: the alternative is a product optimized to hold the twelve percent exactly where they are.
The study cannot prove the companion made anyone lonelier. It shows something nearly as useful: where the students doing worst went when they needed someone, and that when they got there, nothing sent them back.
Sources: The Jed Foundation, “Understanding the Relationships Between AI and Mental Health Among Students” (full report, October 2026; 5,560 students at 21 US schools, October 2025 to April 2026, weighted, generalized mixed additive models adjusted for gender, race/ethnicity, and grade; developed with input from the Digital Wellness Lab at Boston Children’s Hospital; funded by the Nielsen Foundation), as reported by Tech Times (October 27, 2026; use shares, outcome rankings, demographic findings, Kayla Anderson quotes). The Jed Foundation, press release on preliminary findings and Capitol Hill briefing (July 23, 2026). University of Iowa, Center for Social Science Innovation (October 15, 2026; 12 percent of adult AI users use it for companionship). McBain and colleagues, “AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults,” JAMA Pediatrics (June 2026; 19.2 percent, 63.3 percent told no one).
