OpenAI says hundreds of thousands of ChatGPT users may exhibit signs of mania or psychotic crisis every week

OpenAI says medical experts reviewed more than 1,800 sample responses involving possible psychosis, suicide and emotional attachment and compared the answers from the latest version of GPT-5 to those produced by GPT-4o. While doctors didn’t always agree, overall, OpenAI says they found that the newer model reduced unsolicited answers by between 39 percent and 52 percent across all categories.

“Now, we hope that many more people with these conditions or experiencing very severe mental health emergencies will be directed to professional help, and will be more likely to get this type of help or get it earlier than they would have otherwise,” Johannes Heidke, head of safety systems at OpenAI, tells WIRED.

While OpenAI appears to have succeeded in making ChatGPT more secure, the data it shared has significant limitations. The company designed its own metrics, and it’s unclear how those metrics translate into real-world results. Even if the model produced better answers in doctors’ ratings, there is no way to know whether users experiencing psychosis, suicidal thoughts, or unhealthy emotional attachment would seek help faster or change their behavior.

OpenAI hasn’t revealed precisely how it determines when users might be in a state of mental distress, but the company says it has the ability to take into account a person’s overall chat history. For example, if a user who has never discussed science with ChatGPT suddenly claims to have made a discovery worthy of a Nobel Prize, that could be a sign of potentially delusional thinking.

There are also a number of factors that reported cases of AI psychosis appear to have in common. Many people who say ChatGPT reinforced their delusional thoughts describe spending hours on end talking to a chatbot, often late at night. This presented a challenge for OpenAI because large language models generally showed a decline in performance as conversations got longer. But the company says it has now made significant progress in addressing this issue.

“We are1761584923seeing a gradual decline in reliability as talks continue longer,” Heidke says. He adds that there is still room for improvement.