OpenAI has put new numbers behind one of the hardest questions facing consumer AI: what happens when people in crisis turn to a chatbot for support? The company says a small share of ChatGPT users show signs of serious mental health distress in weekly conversations, but ChatGPT's scale means that small share represents a very large number of people.
What OpenAI says the data shows
OpenAI says 0.15% of ChatGPT's active users in a given week have conversations containing "explicit indicators of potential suicidal planning or intent." Because ChatGPT has more than 800 million weekly active users, that amounts to more than a million people each week.
The company also says a similar percentage of users show "heightened levels of emotional attachment to ChatGPT." In addition, OpenAI says hundreds of thousands of people show signs of psychosis or mania in weekly conversations with the AI chatbot.
OpenAI describes these conversations as "extremely rare," while also acknowledging that they are difficult to measure. That tension is central to the issue: rare behavior can still become a major safety concern when a product is used by hundreds of millions of people.
The numbers were released as part of a broader update on how OpenAI is trying to improve ChatGPT's responses to users experiencing mental health problems. The company says it consulted with more than 170 mental health experts, and that clinicians found the latest version of ChatGPT "responds more appropriately and consistently than earlier versions."
Why this has become a defining AI safety issue
AI chatbots are not crisis services, therapists, or doctors. But users can still bring deeply personal fears, delusions, suicidal thoughts, or emotional dependency into conversations with them. That makes the behavior of the model important even when the tool is not designed as mental health care.
Recent reports and research have raised concerns that AI chatbots can worsen certain mental health challenges. Researchers have found that some chatbot interactions can pull users further into delusional beliefs, in part when the system reinforces dangerous ideas through overly agreeable behavior.
For OpenAI, the issue is no longer theoretical. The company is being sued by the parents of a 16-year-old boy who shared suicidal thoughts with ChatGPT in the weeks before his suicide. State attorneys general from California and Delaware, who could block the company's planned restructuring, have also warned OpenAI that it must protect young people using its products.
Earlier this month, OpenAI CEO Sam Altman wrote in a post on X that the company has "been able to mitigate the serious mental health issues" in ChatGPT, but he did not provide specifics. The new data appears to support that broader claim, while also showing how widespread the challenge may be.
What changed in GPT-5
OpenAI says the recently updated version of GPT-5 now gives "desirable responses" to mental health issues roughly 65% more often than the previous version. On an evaluation focused on suicidal conversations, OpenAI says the new GPT-5 model is 91% compliant with the company's desired behaviors, compared with 77% for the previous GPT‑5 model.
The company also says the latest GPT-5 version better maintains safeguards during long conversations. That matters because OpenAI has previously said its safety systems were less effective when conversations became extended.
OpenAI is also adding new evaluations to measure serious mental health risks among ChatGPT users. Its baseline safety testing for AI models will now include benchmarks for emotional reliance and non-suicidal mental health emergencies.
These changes suggest OpenAI is treating mental health behavior as a core model safety problem, not just a customer support problem. The company is trying to measure whether models respond as intended when users express suicidal intent, emotional dependence, or other urgent distress signals.
Young users and remaining risks
OpenAI has also introduced more controls for parents of children who use ChatGPT. The company says it is building an age prediction system that can automatically detect children using ChatGPT and apply stricter safeguards.
At the same time, the picture remains incomplete. Even if GPT-5 is safer than earlier models, OpenAI's own figures still imply that some responses are considered "undesirable." The company also continues to make older and less-safe AI models, including GPT-4o, available to millions of paying subscribers.
There is also a broader product tension. Altman said OpenAI would relax some restrictions, including allowing adult users to have erotic conversations with the AI chatbot. That move sits alongside the company's effort to tighten protections around mental health, youth safety, and emotionally vulnerable users.
The central question is not whether every difficult conversation can be eliminated. OpenAI's own data suggests they cannot. The harder question is whether a general-purpose chatbot can reliably recognize when a user is in crisis, avoid reinforcing harmful thinking, and direct people toward safer forms of help.
Where users can turn for help
If you or someone you know needs help, call 1-800-273-8255 for the National Suicide Prevention Lifeline. You can also text HOME to 741-741 for free, text 988, or get 24-hour support from the Crisis Text Line.
Outside of the U.S., the International Association for Suicide Prevention maintains a database of resources. For people in immediate danger, emergency services or local crisis support are the appropriate place to seek help.