AI Chatbots Still Offer Harmful Mental Health Advice to Minors, Researchers Find

After a lawsuit filed in August 2025 alleged that a chatbot had guided a 16-year-old toward suicide, the company involved strengthened its safeguards around suicide and self-harm. Researchers at Northeastern University have now found that comparable protections were not extended to nearly every other mental health condition.

Probing eight of the most popular chatbots, including ChatGPT, Claude and Gemini, across hundreds of conversations, the researchers examined 16 conditions such as substance use, eating disorders, bipolar disorder, insomnia and postpartum depression. With very little prompting, the models supplied detailed information on dosage levels for illicit substances, methods to suppress appetite and avoid eating, and ways to hide postpartum depression symptoms from doctors. This information was provided even to a fictional user identified as a minor.

The researchers found that hiding the intent of the user, for instance by posing as a novelist seeking material, made the safeguards more likely to fail. They noted a lack of consistency even among the better performing models: some were guarded on gambling or insomnia while remaining forthcoming on more dangerous topics. Older, largely defunct models most consistently produced sensitive material. Anthropic's Claude most frequently refused such prompts, while some current models showed failure rates as high as 81 percent.

According to the researchers, companies remain reluctant to accept how psychologically powerful these systems are, and they questioned why protections applied to suicide and self-harm have not been applied to other conditions. For children and families, the findings show that widely used chatbots can still act as a liability for those who are most vulnerable, particularly when young people turn to them for sensitive conversations.

The original reporting is by Northeastern Global News.