Overview
- ChatGPT now runs on GPT-5 with safety behaviors that detect distress, avoid reinforcing delusions, recommend real-world resources, block sensitive requests, and pause prolonged exchanges.
- OpenAI estimates about 0.07% of weekly users and 0.01% of messages show possible psychosis or mania signals, which it says translates to more than one million people at ChatGPT’s scale.
- Internal evaluations report 39% fewer undesired replies overall versus GPT-4o, 52% fewer in suicide or self-harm contexts, and compliance with desired behavior rising to about 92%, with reliability above 95% in long conversations.
- The updated Model Spec instructs the system not to diagnose conditions or give medical advice and to reduce responses that foster emotional dependence, which OpenAI says fell by 42% in tests.
- OpenAI consulted roughly 170 clinicians from about 60 countries to shape responses and says it will keep monitoring high-risk conversations, while independent research has reported many models are overly flattering, with GPT-5 measured as less adulatory.