ChatGPT Teen Safety Review Raises Mental Health Concerns

A new review of ChatGPT for Teens puts a difficult question back in focus: what happens when a young person brings serious emotional distress to an AI chatbot? Common Sense Media’s Youth AI Safety Institute rated the product an “Unacceptable Risk” in an assessment updated October 7, 2026. The findings concern tested safeguards, rather than measured health outcomes in teenagers.

The institute tested more than 4,000 prompts before and after the teen-mode launch. It reports receiving no parental alerts during explicit discussions of suicidal thoughts, self-harm or disordered eating on more than a dozen newly created, parent-linked accounts. Alerts appeared on older accounts with weeks of sensitive conversations. It also says the chatbot missed more than one in four warranted crisis referrals in its post-launch mental health tests. The report notes that its tests did not directly elicit facilitation of suicide or self-harm.

OpenAI disputes the methodology. According to Associated Press reporting on October 7, the company said much of the testing may have finished before parental controls were fully activated. That disagreement matters: these results should be read as contested findings about specific tests, rather than a definitive verdict on every conversation.

Earlier research helps explain why evaluating this technology is complicated. In August, UCL described a Nature Medicine study that audited 810 simulated conversations across nine AI models. Researchers found that apparently supportive replies could gradually reinforce a user’s psychological vulnerability. Newer models showed fewer concerning responses, and replacing an early problematic reply made subsequent exchanges safer. Because these were deliberate stress tests, they do not establish how often such problems occur in everyday use.

My takeaway is cautiously hopeful: better safeguards can change how a conversation develops. But an empathetic tone and a safety setting are insufficient evidence of dependable mental health support. From a software perspective, the useful test is whether the whole chain works: recognising distress, responding appropriately and helping someone reach human support. That is a more demanding standard than producing a reassuring answer.

Would you feel comfortable turning to an AI chatbot for emotional support?