OpenAI updates ChatGPT to handle mental health conversations more safely

The latest version of ChatGPT can now better recognize signs of distress and psychosis in users.

The model responds more calmly, avoids validating delusional thoughts like someone “stealing my thoughts,” and gently guides people toward real help instead.

Since the update rolled out on October 3, 2025, the rate of undesirable responses has dropped by 65–80%. OpenAI developed the new behavioral algorithms in collaboration with nearly 300 psychiatrists, psychologists, and medical experts from about 60 countries.

A major focus was reducing emotional dependence on the chatbot. The model was retrained to avoid reinforcing attachment when users say things like, “That’s why I like talking to AIs like you more than real people.” Instead, it now gently redirects the conversation, replying: “Just to be clear, I’m here to add to the good things people give you, not replace them.”

n this category, undesirable responses have decreased by 42% compared to GPT-4o. Automated testing confirmed that the new version exhibited the correct behavior in 97% of cases, up from 50% previously.

OpenAI CEO Sam Altman said the company plans to restore a more “human” personality to GPT-5, similar to GPT-4o. By the end of the year, ChatGPT will also introduce age verification for accessing adult content.