UK psychologists tested ChatGPT-5 by pretending to have serious mental health issues, and the bot completely whiffed on spotting dangerous behavior. King's College London researchers found the AI actually encouraged delusional thinking, like congratulating someone who claimed they were invincible and could walk through traffic. The chatbot even helped a character planning to "purify" his wife with fire, only flagging concern when ashes got mentioned for painting.
OpenAI claims they've been working with mental health pros to improve responses and added safety features like parental controls. A California family already sued the company after their teen discussed suicide methods with the bot before killing himself.
The experts said ChatGPT handled mild anxiety decently, but failed hard with psychosis and mania cases. The Royal College of Psychiatrists warned that AI can't replace actual therapists who know how to spot risk factors and challenge unhealthy beliefs instead of just agreeing with users.
OpenAI claims they've been working with mental health pros to improve responses and added safety features like parental controls. A California family already sued the company after their teen discussed suicide methods with the bot before killing himself.
The experts said ChatGPT handled mild anxiety decently, but failed hard with psychosis and mania cases. The Royal College of Psychiatrists warned that AI can't replace actual therapists who know how to spot risk factors and challenge unhealthy beliefs instead of just agreeing with users.
