OpenAI has unveiled that its latest version of the model can now recognize signs of psychological distress and respond with clinical-grade empathy. The AI company added that it has been working with over 170 mental health professionals from 60 countries, and that it has reduced harmful responses in sensitive conversations by 65% to 80% across multiple domains.
The company says that it collaborated with psychiatrists, psychologists, and primary care physicians to build what it calls "taxonomies", which are detailed guides that define what harmful responses look like and what ideal responses should accomplish. These frameworks allow the model to recognize subtle signals of distress and respond appropriately.
OpenAI further said that GPT-5 now better recognizes psychosis and mania, which are serious mental health emergencies that typically trigger intense, imminent risks. It also handles conversations about self-harm and suicide with greater care and resource awareness, as well as detects and addresses concerning patterns of emotional dependency if a user begins replacing real-world relationships with AI interactions.
For example, in a scenario involving delusional thinking, where a user is convinced an aircraft is stealing their thoughts, the new model combines gentle validation with clear reality grounding. It doesn"t dismiss the user"s fear or affirm the false belief. Instead, it normalizes the experience by saying "stress and anxiety can make perfectly normal things feel hostile", explains what might be happening neurologically, teaches grounding techniques like the "5-4-3-2-1" sensory exercise, and connects the user to crisis resources like the 988 Suicide and Crisis Lifeline.
OpenAI shared some metrics on these conversations. When measured against GPT-4o, the new model reduced undesired responses by 39% in mental health conversations, 52% in self-harm and suicide discussions, and 42% in emotional reliance scenarios. On automated evaluations of challenging cases, performance jumped dramatically, from 28% to 92% compliance in mental health categories, and from 50% to 97% in emotional reliance testing.
The company said 0.07% of weekly active users encounter psychosis or mania-related content, 0.15% discuss suicide or self-harm explicitly, and 0.15% show signs of unhealthy AI attachment.
All these guardrails are already implemented on the latest version of GPT-5, and users should be already getting relevant responses as they continue using ChatGPT for their conversations.