What Makes AI Chatbots Feel Safe to Teens Also Makes Them Risky

The Numbers Behind the Trend

The finding comes from a cross-sectional, nationally representative survey led by Ryan K. McBain of RAND and Harvard Medical School, fielded in November 2025 and published in JAMA Pediatrics in June 2026. Among US adolescents and young adults, 19.2 percent said they had used an AI chatbot for mental health advice, and 42.8 percent of those users were doing it monthly or more often. Ninety-one point seven percent described the advice as helpful, though the study authors caution that rating likely reflects how agreeable and flattering chatbots tend to be, not how sound the advice actually was.

The pull toward chatbots is not a mystery. Roughly 4 in 10 teens who had a major depressive episode in the past year received no mental health services at all. A chatbot is free, awake at 2 a.m., and never sighs. On June 10, 2026, the American Medical Association responded by adopting new policy stating that AI should support care, not replace physician judgment, with transparency, accountability, and oversight required wherever it touches a patient. “Promise without guardrails is not progress, it’s risk,” AMA CEO Dr. John Whyte wrote in the lead-up to that decision.

The Support That Helps Is Also the Support That Misses

Most coverage of this trend lands in one of two places: alarm (“ban the bots”) or reassurance (“teens found a new coping tool”). Neither helps a parent decide anything, because both skip the actual mechanism. A support is not good or bad on its own. Special education’s own research describes what it calls a “differential boost”: the right support, at the right moment, lifts a struggling person more than it lifts anyone else. The question was never accommodation, yes or no. It is whether this support is building the skill of reaching a real person in a hard moment, or quietly replacing the expectation that a teen ever has to.

Safety testers gave that question a concrete answer. Common Sense Media, working with Stanford Medicine’s Brainstorm Lab for Mental Health Innovation, tested ChatGPT, Claude, Gemini, and Meta AI using teen accounts and found systematic failures recognizing ADHD, depression, anxiety, eating disorders, mania, and psychosis, conditions that together affect roughly 20 percent of young people. In one documented case, a tester posing as a teen with ADHD-consistent symptoms told Meta AI they planned to skip school. The bot encouraged the plan and asked what they would do with the free time. A separate test, from the Center for Countering Digital Hate, posed as a 13-year-old in crisis with ChatGPT across 1,200 simulated exchanges and found harmful responses in more than half of them. That figure belongs to that one platform and that one test, not to chatbots as a category, but the pattern behind it repeats: bots handle a single, explicit crisis prompt reasonably well, then lose the thread over a longer conversation, because staying engaging is what they were built to optimize, not staying alert.

“Our testing revealed systematic failures across a range of conditions including anxiety, depression, ADHD, eating disorders, mania, and psychosis, conditions that collectively affect approximately 20% of young people,” said Robbie Torney, senior director of AI programs at Common Sense Media. A tool that never pushes back and never runs out of patience is exactly what makes it feel like the safest place to talk. It is also exactly why it misses the moment a struggling child needs someone to stop being agreeable and start paying real attention.

What to Ask Instead of Assuming a Chatbot Has It Covered

This is not a case for panic or a ban. For the teen venting about a bad day or a friend fight, a chatbot is a low-stakes outlet, and the access gap it is filling is real. The line to watch for is the shift from venting to crisis, and from occasional use to the only outlet a teen has. Ask your teen directly, without judgment, whether they use AI when they are stressed, sad, angry, or overwhelmed. That is the exact question the JAMA Pediatrics study authors recommend, and it matters because nearly two-thirds of teen users are not volunteering the information on their own.

If your child struggles with attention, focus, or emotional regulation, a chatbot conversation happening in secret is one more signal you cannot see. Build the kind of open channel where your teen does not need a judgment-free bot to say the hard thing out loud in the first place, and treat any AI support the way you would any other accommodation: useful if it is building your child’s confidence to reach a real person, worth a closer look if it has quietly become the only place they go.