In a blog post, OpenAI said that 0.07 per cent of its 800 million weekly users show indications of serious mental health emergencies, amounting to roughly 560,000 people.

An additional 1.2 million users — or 0.15 per cent — send messages containing what the company calls "explicit indicators of potential suicidal planning or intent".

The company also identified over one million weekly cases of what it terms “exclusive attachment” to the chatbot — a pattern where users form strong emotional bonds with ChatGPT that can interfere with their real-world relationships, well-being, or responsibilities.

The findings have intensified concerns among mental health experts and policymakers about the potential psychological risks of prolonged chatbot use.

Dr. Hamilton Morrin, a psychiatrist at King's College London, said: "It’s encouraging to see companies like OpenAI working with clinicians and researchers to improve model safety.

"But the problem is likely far from solved."

In response, OpenAI has convened a panel of more than 170 mental health experts to help refine how ChatGPT identifies and responds to at-risk users. The company says it has retrained its newest GPT-5 model, achieving 91 per cent compliance with its desired behavior in sensitive interactions — up from 77 per cent in prior versions.

Still, experts caution that these numbers only scratch the surface

Dr. Thomas Pollak, a neuropsychiatrist at South London and Maudsley NHS Foundation Trust, said: "Even a small percentage represents a very large number of people. These figures should be taken seriously, even if interpreted cautiously."

An OpenAI spokesperson added that measuring such interactions remains "extremely difficult" and that results could shift as the company’s understanding deepens.