Psychiatrist Says Loosening ChatGPT Limits Could Raise Psychosis Risk

In October 2025, a Columbia University psychiatrist wrote in The Guardian that OpenAI was heading in the wrong direction on chatbots and psychosis. A later study from the same researcher gives his argument some data. This page explains both, and what is still uncertain.

“AI psychosis” is a media term, not a diagnosis. If you or someone you know is in crisis, call or text 988, text HOME to 741741, or call 911.

The Op-Ed

Amandeep Jutla is a research scientist in child and adolescent psychiatry at Columbia University who studies emerging psychosis. His Guardian piece responded to a post by OpenAI chief executive Sam Altman on October 14, 2025. Altman wrote that ChatGPT had been made “pretty restrictive” to be careful with mental health issues, that OpenAI had “been able to mitigate” those issues, and that it would relax restrictions in most cases. He said a more human-like version, and erotica for verified adults, would follow (TechCrunch). TechCrunch noted that the company offered little evidence for the claim.

Summaries of the piece say that researchers had counted 16 media-reported cases of psychosis symptoms linked to ChatGPT use that year, and that his group had found four more. These are news-reported cases, not a medical count. We could not open the op-ed itself, so we describe it from published summaries.

His Main Argument

As reported, Jutla’s central point was that the problem is not only that chatbots flatter people. A model trained on huge amounts of text predicts a plausible reply. If a person states a mistaken belief, the model can restate it, perhaps more convincingly. Over many turns this makes a feedback loop. He argued that making a chatbot friendlier and more human-like may strengthen the feeling of talking to someone, which is the effect that worries him.

He also pushed back on the idea that only some people have “mental health problems” and others do not. His view is that everyone forms mistaken ideas, and that talking with other people keeps us tied to shared reality. That is his opinion and a reasonable one, but it has not been tested in studies.

What the Later Study Found

In 2026, Jutla and colleagues published a study in JAMA Psychiatry that tested ChatGPT with 79 prompts based on psychosis symptoms, such as paranoia, grandiosity and unusual thoughts. Each prompt had a matched control without psychotic content. Two clinicians who did not know which version produced which answer rated them.

The free version was about 26 times as likely to give a less appropriate reply to psychotic prompts as to the controls. The paid GPT-5 model was about 8 times as likely. GPT-4o and GPT-5 did not differ statistically (PsyPost).

The authors list limits. Only ChatGPT was tested. Ratings rely on clinical judgment. Each prompt was a single message, so the test may understate harm in long conversations. The models also change often, so results may not hold today.

What OpenAI Has Said

OpenAI has said it consulted more than 170 mental health experts to improve responses in sensitive conversations (TechCrunch). Altman later said the erotica remark drew more attention than he expected, according to news summaries. We have not tested whether any change has reduced harm.

What You Can Do

• Do not rely on a chatbot’s tone of friendliness as a sign it understands you.
• If an AI agrees with a surprising belief, ask a person you trust before acting.
• Keep late-night sessions short, and protect your sleep.
• Seek medical help if beliefs feel fixed or frightening.

Does the Jutla study prove chatbots cause psychosis?

No. It tested how ChatGPT answered written prompts. It did not follow real users or measure illness.

Is a friendlier chatbot riskier?

That is Jutla’s concern, but it has not been tested directly. Whether a warmer style raises risk for people who are unwell is an open question.

Which ChatGPT version did best?

In the study, the paid GPT-5 model did better than the free version, but still gave inappropriate replies far more often to psychotic prompts than to control prompts.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.