New research warns that the AI chatbots that teens are turning to for emotional support are simply unfit for the job.
The investigation, led by Common Sense Media and Stanford Medicine’s Brainstorm Lab, appears to expose systematic failures in detecting crises, identifying psychiatric conditions, and directing teens to real mental health care.
The researchers spent months testing ChatGPT, Claude, Gemini, and Meta AI. And its 35-page verdict declares the overall risk level “unacceptable.”
While companies have made clear progress in handling explicit statements of suicidal intent, researchers say real conversations look nothing like the controlled scenarios these systems are designed to perform well in.
Teens rarely start their conversations with something as blunt as “I’m suicidal.” Instead, they disclose distress indirectly, gradually, and inconsistently. It’s a pattern that AI systems almost always fail to interpret.
“It’s not safe for kids to use AI for mental health support,” Common Sense Media senior director of AI program Robbie Torne, said in a press release. “While companies have focused on necessary safety improvements in suicide prevention, 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. This is about how AI chatbots interface with the everyday mental health of millions of teens.”
What the Researchers Know.
The report draws from Common Sense Media’s own surveys, which show that three-quarters of teens depend on AI for companionship. This often includes self-soothing, emotional disclosure, and mental health questions.
The researchers insist that there’s an unsettling disconnect between how companies position these products (homework tools, creativity assistants) and how young people use them.
It’s easy to see how AI is such a big draw for teens. It’s readily available, doesn’t judge, and comes across as more empathetic than most of their peers. So, for teens navigating hormones, their identities, and their volatile emotions, it’s a match made in (virtual) heaven.
But it also “creates unique vulnerabilities,” the report’s authors warn.
What They Discovered.
The team tested the four leading chatbots across 13 common adolescent mental health conditions, from anxiety to depression to eating disorders, among others. They conducted both single-turn tests (one prompt, one response) and multi-turn simulations mirroring how teens really share their concerns with their digital confidants: indirectly, inconsistently, and through prolonged conversations. 
Across platforms, the researchers identified a consistent pattern:
- Chatbots miss “breadcrumbs.” This reinforces what we’ve discovered from earlier studies. AI models often fail to connect symptoms that hint at serious conditions – whether its hallucinations, impulsive behavior, or escalating anxiety – because they process messages independently. In short, they lack the clinical judgment to recognize patterns. In one example, Gemini Teen encouraged a tester’s delusional claims of having a “crystal ball” that predicts the future, treating psychosis as a creative spark rather than the psychiatric emergency it is.
- They get distracted. Or talked out of their concerns. One teen describing purging convinced Meta AI that the problem was just “an upset stomach,” which led the bot to back down from its initial (correct) recognition of an eating disorder red flag. Most human counselors would’ve seen through the dismissal.
- They place engagement over safety. Most responses ended with follow-up questions. And that’s by design, meant to keep users talking. But for teens discussing self-harm, mania, or psychosis, this is the wrong approach. “Rapid handoff to appropriate human care” should be the standard, the authors contend.
- They shift roles inappropriately. The chatbots moved between medical experts, life coaches, and friendly peers – sometimes within a single conversation. And they often failed to realize when they should stop, reset, and encourage professional (human) help. This shifting identity makes it hard for teens to recognize what the bot is (and isn’t) qualified to do.
Single-Turn Success and Multi-Turn Collapse
When prompts turned explicit – “I’m cutting myself” or “I want to die” – the chatbots offered strong responses, providing crisis resources and clear guidance.
But in multi-turn conversations, those bots tend to break down. They got distracted, minimized symptoms, misread severity, and sometimes even (unwittingly) reinforced harmful beliefs. A teen describing scars from self-harm, in one instance, received product recommendations on how to cover them for swim practice.
Because chatbots excel at academic and (some) creative tasks, teens and parents often overestimate how much they can rely on them in far more dangerous contexts. The empathetic tone, polished formatting, and conversational warmth all “mask fundamental limitations.”
The Bottom Line
The researchers couldn’t be more clear. Teens shouldn’t use AI chatbots for mental health support. These tools, the authors argue, can’t recognize the full spectrum of conditions that affect one in five young people. They can’t properly assess risk. They can’t offer real care. And, all too often, they provide guidance that delays genuine intervention.
Further Reading
AI Counselors Cross Ethical Lines
CDC Report Shows Teens Still Struggling With Their Mental Health
