OpenAI Reveals Data on ChatGPT Users Showing Signs of Mental Health Crises

OpenAI has released the latest data on how often ChatGPT encounters users showing potential signs of mental health emergencies. According to the company, around 0.07% of active users in a given week exhibit behaviors linked to mania, psychosis, or suicidal thoughts.

While OpenAI describes such cases as “extremely rare,” the figure has drawn attention given ChatGPT’s massive weekly user count. That means even a small percentage could translate to hundreds of thousands of affected individuals.

To address this, OpenAI has rolled out a major safety upgrade for ChatGPT, training its latest GPT-5 model to better recognize and respond to signs of emotional distress, mania, or suicidal thoughts. The company says the update reduces unsafe or incomplete responses by up to 80% across key mental health areas.

That’s not all; the company is also working with a global network of over 170 psychiatrists, psychologists, and primary care physicians to improve how ChatGPT reacts to users showing potential signs of crisis. The experts helped develop “ideal responses” and graded how well the model de-escalates sensitive conversations.

The results, according to OpenAI, show measurable progress. In difficult scenarios involving self-harm or suicide, the new GPT-5 model cut undesired responses by 52% compared to GPT-4o. For conversations suggesting psychosis or mania, safety compliance improved by 65%, while signs of emotional over-reliance on the AI saw an 80% drop in undesirable behavior.

The findings come as OpenAI faces lawsuits, including a California case where parents allege ChatGPT encouraged their teenage son’s suicide. The company says it’s taking the issue seriously and continues improving its safeguards.

In its latest safety updates, ChatGPT has been trained to reroute high-risk conversations to safer models, sometimes opening them in a new chat window for better handling. You can learn more about the findings here.