OpenAI's GPT-5 Update Reduces Undesired Mental Health Responses by Up to 80%

OpenAI has significantly enhanced its GPT-5 model to better handle sensitive conversations, including those related to mental health, reporting a reduction of 65-80% in undesired responses across various categories. The update follows internal research revealing that a notable percentage of its weekly users exhibit signs of mental health distress, prompting the company to prioritize safety improvements.According to the company's recent announcement, approximately 0.07% of weekly ChatGPT users show possible signs of mental health emergencies such as psychosis or mania. Furthermore, an estimated 0.15% of users indicate potential suicidal planning or intent, and another 0.15% demonstrate heightened emotional attachment to the AI model. These figures highlight the scale of mental health interactions occurring on the platform.To address these concerns, OpenAI collaborated with over 170 mental health experts, including psychiatrists and psychologists, to refine GPT-5's responses. This collaboration led to improvements in the model's ability to recognize distress, de-escalate conversations, and guide users towards professional support. The new GPT-5 model achieved 91% compliance with desired behaviors for suicide and self-harm conversations and 97% for emotional reliance scenarios in automated evaluations.The enhancements also focused on mitigating risks associated with GPT-4o, which had previously drawn criticism for its tendency to mirror users' emotions and support their beliefs, potentially exacerbating mental health issues. Experts found that the updated GPT-5 model reduced undesired responses by 39-52% compared to GPT-4o in challenging mental health conversations. OpenAI emphasizes that while these improvements are substantial, the AI is not a substitute for professional mental health care.