OpenAI says the medical experts reviewed more than 1,800 model responses involving potential psychosis, suicide, and emotional attachment and compared the answers from the latest version of GPT-5 to those produced by GPT-4o. While the clinicians did not always agree, overall, OpenAI says they found the newer model reduced undesired answers between 39 percent and 52 percent across all of the categories.
“Now, hopefully a lot more people who are struggling with these conditions or who are experiencing these very intense mental health emergencies might be able to be directed to professional help, and be more likely to get this kind of help or get it earlier than they would have otherwise,” Johannes Heidecke, OpenAI’s safety systems lead, tells WIRED.
While OpenAI appears to have succeeded in making ChatGPT safer, the data it shared has significant limitations. The company designed its own benchmarks, and it’s unclear how these metrics translate into real-world outcomes. Even if the model produced better answers in the doctor evaluations, there is no way to know whether users experiencing psychosis, suicidal thoughts, or unhealthy emotional attachment will actually seek help faster or change their behavior.
OpenAI hasn’t disclosed precisely how it identifies when users may be in mental distress, but the company says that it has the ability to take into account the 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 possible delusional thinking.
There are also a number of factors that reported cases of AI psychosis appear to share. Many people who say ChatGPT reinforced their delusional thoughts describe spending hours at a time talking to the chatbot, often late at night. That posed a challenge for OpenAI because large language models generally have been shown to degrade in performance as conversations get longer. But the company says it has now made significant progress addressing the issue.
“We 1761584671 see much less of this gradual decline in reliability as conversations go on longer,” says Heidecke. He adds that there is still room for improvement.
