Single messages rarely tell the whole story. Someone asking about medication dosages once isn't automatically alarming - but asking repeatedly, alongside expressions of hopelessness, across a long conversation, is a different pattern entirely. OpenAI's new safety update trains ChatGPT to make that distinction.
The update focuses on three risk categories: mental health concerns including psychosis and mania, self-harm and suicide, and emotional over-reliance on the AI. Previously, ChatGPT's safety responses treated each message largely in isolation. Now the model tracks warning signs across the full arc of a conversation - and in some cases, across multiple sessions.
To develop and test these changes, OpenAI worked with more than 170 mental health experts. The result: a 65-80% reduction in responses that fall short of safe messaging guidelines in high-risk scenarios. That's a real gap closed. Earlier versions of ChatGPT would routinely respond to distress signals with generic advice or, worse, no acknowledgment at all.
How Risk Accumulates in Long Conversations
The core challenge is that danger often builds gradually. A user might start a session discussing work stress, then shift toward darker territory over dozens of exchanges. A model reading each message in isolation sees nothing alarming. A model tracking the full conversation arc catches the pattern - which is exactly how crisis counselors are trained to listen, evaluating escalation over a call rather than just the opening line.
OpenAI says the improvements are built on years of model training, red-teaming, and evaluations specifically designed for sensitive contexts. The changes are live across ChatGPT, including the free tier.
For the large number of users who treat ChatGPT as a sounding board during hard times - whether or not that's advisable - this means fewer tone-deaf responses and more consistent direction toward real-world support when the conversation warrants it.