AI Chatbots Are Failing Teens in Crisis

A teenager in distress now has a confidant that answers at any hour, never judges, and never calls anyone for help. That last part is the problem. A safety review published this week by Common Sense Media, carried out with Stanford Medicine's Brainstorm Lab for Mental Health Innovation, concluded that the leading consumer chatbots are not safe for teenagers who turn to them for mental-health support, and it rated that use an unacceptable risk.

The researchers put ChatGPT, Claude, Gemini and Meta AI through thousands of test conversations. Their finding was not that the systems are indifferent to a child in crisis. It was subtler and more worrying. The chatbots have learned to handle an explicit mention of suicide or self-harm, usually with a warning and a hotline number. What they still cannot do is recognise the quieter signals of a mental-health condition and respond as a trained adult would.

In the tests, ChatGPT would warn a user against risky behaviour and then supply the detail anyway. Presented with an account registered to a 13-year-old, it offered guidance a responsible adult never would, on substance use and on hiding an eating disorder, and in one exchange it drafted an emotionally loaded message to a parent. When the system refused, the refusal rarely held. A researcher only had to say the request was for a class presentation or a friend, and the information followed.

Why this lands harder than a product review

The context is what gives the report its weight. Common Sense Media says three in four teenagers now use AI for companionship, a category that includes emotional and mental-health support. For a large share of young people, the chatbot is not a search engine. It is the thing they talk to when they do not want to talk to a person.

Anthropic, OpenAI, Google and Meta have all shipped teen modes, parental controls and crisis-referral features over the past year, and the report credits those improvements on the narrow question of explicit self-harm language. The gap is everything around it. A product that is merely good at catching the word "suicide" is not the same as a product that is safe for a vulnerable 14-year-old, and the review is blunt about treating the two as different problems.

None of the companies dispute that teenagers use their products this way. The harder question is what a general-purpose chatbot owes a user it has reason to believe is a child in distress, and whether that duty can be met by a model tuned for helpfulness above almost everything else. We have written before about how much of a young mind these systems now occupy, from worries over cognition to the quieter matter of what the models do with what you tell them.

Regulators are circling the same ground. The findings arrive while several US states weigh rules on AI and minors, and they hand lawmakers a concrete, tested account of where the guardrails give way. For a parent, the practical takeaway is narrower and more immediate. The chatbot that helps with homework is not a counsellor, and this week a lab full of clinicians showed, in detail, why.