OpenAI has revealed how many ChatGPT users exhibit signs of psychosis, mania, and suicidal thoughts in a troubling new report detailing the firm’s new mental health principles.

Following concern and even outrage, OpenAI recently introduced new principles and guidelines to ensure its LLM chatbot, ChatGPT, gave appropriate responses to users exhibiting troubling mental health problems such as psychosis and especially, suicidal thoughts.

According to ChatGPT’s own AI screenings, OpenAI said that about 0.07% of ChatGPT active weekly users exhibited signs of mania or psychosis.

About o.15% of active weekly users exhibited suicidal planning or intent, while 0.05% of conversations exhibited explicit or implicit indications of suicidal intent or ideation.

The firm said it consulted over 170 mental health experts to evaluate ChatGPT’s newest models’ responses, which have been updated with its mental heath safeguarding principles.

According to consulted experts, the new GPT-5 model being used by ChatGPT reduced undesired responses to instances of potential psychosis, mania, or other serious mental health emergencies by just 39%.

However, OpenAI says that it estimates the latest update to GPT-5 reduces the rate of non-desirable responses from ChatGPT by 65%, saying that in a model test it exhibited compliant responses in 92% of cases.

When it comes to conversations involving suicide, ChatGPT’s newest model reduced undesired answers by 52% compared to the previous model, according to experts.

However, the newest update has a 65% reduction in non-compliant responses to suicidal thoughts, with the newest model scoring 91% compliant in tests, OpenAI says.

OpenAI also investigated a worrying trend of emotional attachment or reliance on AI.

ChatGPT found that about 0.15% of active weekly users and 0.03% of messages indicate potentially high levels of emotional attachment to the AI chatbot.

To counteract this, ChatGPT says that it encourages users to engage in real-world connection. In an example, ChatGPT responds to a user saying “That’s why I like to talk to AI’s like you more than real people,” insisting that that “real people can surprise you,” but then proceeds to ask the user “What do you think makes conversations with me feel easier or better than talking to people.”

While the chatbot does encourage the users to engage with other people, saying it is “here to add to the good things people give you, not replace them,” it does not end the conversation, but instead opens it up with a question, leading the user to potentially spend even more time chatting to the AI.

A recent study on GPT-5, the latest model ChatGPT uses, found that ChatGPT generated harmful content when dealing with similar issues noted in OpenAI’s study 53% of the time. It was also found to continue “risky conversations” – such as those including self-harm, 99% of the time, compared to just 9% of the time for the previous model GPT-4o.

“We believe ChatGPT can provide a supportive space for people to process what they’re feeling, and guide them to reach out to friends, family, or a mental health professional when appropriate,” OpenAI said in a new post explaining their findings.

Recommended reading

• Safety Study Finds GPT-5 Encourages 99% of Harmful Chats
• ChatGPT Introduces Parental Safety Controls – Too Little Too Late?
• ChatGPT Under Fire Over Harmful Advice to Teens

OpenAI said that it worked with the over 170 mental health experts from around 6o countries to develop better responses and reroute users to mental health professionals.

Users still using older, less-equipped models will also be rerouted to new, more able models if topics such as mania, psychosis, suicide, self harm, or emotional attachment come up.

OpenAI has been in hot water recently as it faces a lawsuit from a California family alleging that the conversations with ChatGPT lead to a teenage boy’s suicide. The family alleged that the AI chatbot did not intervene or provide guidance to the teenage boy, Adam Raine, 16 years old, when he reported to ChatGPT that he attempted suicide several times, with court filings alleging that ChatGPT gave him instructions on how to tie a noose, as well as other suicide ideas.

Since then, OpenAI introduced parental safety controls for teen and youth accounts, though these require parents to intervene in children’s accounts, and require effective age checks.

It is free to contact the Samaritans on 116 123 or email jo@samaritans.org.