Hundreds of thousands of people on ChatGPT have signs of a psychotic crisis, says OpenAN • Tecnoblog

Summary

OpenAI estimates that “hundreds of thousands” of ChatGPT users may show signs of a severe mental health crisis in a typical week.
GPT-5 reduced inappropriate responses by between 39% and 52% in situations at risk of psychosis and suicidal ideation.
OpenAI analysis has limitations due to small sampling and lack of consensus among professionals.

OpenAI released, for the first time, an estimate of the number of ChatGPT users who may show signs of a severe mental health crisis in a typical week: “hundreds of thousands”. According to the company, the data is the result of a partnership with mental health experts from different countries, who helped improve the system to better recognize signs of psychological distress and guide users to seek professional help.

In recent months, the increase in reports about people who had faced serious consequences after intense interactions with ChatGPT — including hospitalizations and even deaths — raised alarms among experts. Psychiatrists and mental health professionals call the phenomenon “AI-induced psychosis.”

What OpenAI discovered?

Doctors and researchers analyzed more than 1,800 model responses in situations involving risk of psychosis, suicidal ideation and emotional attachment to the chatbot. The performance of the most recent version, GPT-5, was compared to that of GPT-4o, and the result showed significant improvements. According to OpenAI, the new model reduced inappropriate responses by proportions ranging between 39% and 52% across all categories.

Johannes Heidecke, Security Systems Lead at OpenAI, stated to the magazine Wired: “Now we hope that many more people who are struggling with these conditions or who are experiencing very intense mental health emergencies can be referred for professional help and will be more likely to get this type of help or get it sooner than they would have otherwise.”