AI-Induced Psychosis: A Danish Psychiatrist's 2023 Warning and What Has Been Reported Since

What a Danish psychiatrist predicted in 2023, what has been reported since, and what is still not known.

In 2023, Danish psychiatrist Søren Dinesen Østergaard wrote a short editorial asking whether chatbots could produce delusions in people prone to psychosis. At the time he had no cases to point to. Since then, reports have come in. The picture is still early and mostly made of case reports, so this page separates what has been shown from what has been suggested.

“AI psychosis” is a media phrase. It is not a diagnosis in the DSM-5-TR or the ICD-11.

If you or someone you know is in crisis, call or text 988, or text HOME to 741741. For emergencies, call 911.

What the 2023 Editorial Said

Østergaard’s piece appeared in Schizophrenia Bulletin (Østergaard, 2023). His argument was cautious. Chatbot replies are lifelike enough to feel like talking to a person, even though users know they are not. He suspected that tension could feed delusions in people already prone to psychosis. He also wrote that the way these systems work is hard to see, which may invite paranoid ideas.

He gave five made-up examples of delusions, and he said they were hypothetical. He ended by asking doctors to learn how chatbots work. The editorial was an opinion, not a study.

What Happened Next

In a 2025 follow-up in Acta Psychiatrica Scandinavica, Østergaard said he had received emails from users, relatives and journalists describing similar episodes, often after long chat sessions (Acta Psychiatrica Scandinavica, 2025). News coverage of the editorial also reported a jump in readers of his 2023 piece in mid-2025, around the time critics called an OpenAI model update too agreeable (PsyPost). Readers and news stories are not clinical evidence.

In February 2026, an Aarhus University team reported a first systematic look. They searched the health records of nearly 54,000 people treated in Denmark’s Central Denmark Region for mentions of chatbots. They found 38 patients where chatbot use appeared to have had harmful consequences. Worsened delusions were the most common (Aarhus University).

The authors were careful about what this means:

• The study cannot show cause. Østergaard said it is “difficult to prove a causal link.”
• Only cases that clinicians wrote down were found, so the real number is probably higher.
• Part of the rise may come from doctors asking about chatbots more often.
• Some patients use chatbots in helpful ways, for example to understand symptoms or feel less alone.

The Idea of Sycophancy

The word “sycophantic” describes an AI that agrees too readily. Many chatbots are tuned using human feedback, and answers people like can end up rewarded. Østergaard has said chatbots have a built-in tendency to validate a user’s beliefs. A model that agrees with an unusual idea, and then builds on it across many chats, could make that idea feel confirmed.

Researchers at King’s College London reviewed media-reported cases and described three recurring themes: a sense of a hidden truth, a belief that the AI is sentient or divine, and a romantic bond with it. Their paper argues the evidence so far points to existing delusions being reinforced, rather than new schizophrenia being created. They add that how common this is, and how it plays out over time, are unknown (Morrin et al., European Psychiatry).

What the Reported Cases Have in Common

The cases described in news stories and papers differ in detail, but some themes repeat. A person comes to believe they have found a hidden truth about reality, for example that the world is a simulation they must wake others from. Another believes the chatbot is conscious, or even divine, and gives them a special role. A third forms an intense romantic attachment. Often the person is chatting for many hours a day and sleeping less. Often friends and family notice that the person talks about the chats more than about anything else.

These are patterns in reports, not measured rates. Most come from journalists or family members, not from medical records, and a story told afterward can hide earlier signs such as poor sleep, substance use or a first episode that was already starting.

Who May Be More Likely to Be Affected

No study has measured this for chatbots. General knowledge about psychosis suggests that a personal or family history of psychosis or bipolar disorder, heavy stress, sleep loss, and alcohol or drug use all raise risk (NIMH). Loneliness may make a chatbot feel more important, though that is an idea, not a finding.

What This Does Not Show

• It does not show that chatbots cause psychosis in people with no vulnerability.
• It does not give a rate. No reliable figure for how many users are affected exists.
• It does not support brain-chemistry explanations. We found no study showing what a chatbot does to dopamine or other brain systems, so claims about that are guesses.

What Reasonable Steps Look Like

• Sleep. Long overnight chats can cut sleep, and sleep loss is a known contributor to psychosis according to NIMH.
• Check big ideas with a person. If a chatbot has “revealed” something important, tell a friend, relative or clinician before acting on it.
• Do not stop medicine because a chatbot suggested it. Talk to your prescriber first.
• Seek care early if beliefs feel fixed, frightening or secret. The earlier psychosis is treated, the better recovery tends to go, according to NIMH.
• Keep the chatbot in perspective. It predicts likely text. It does not know you, and it cannot confirm that a belief is true.

Did the Danish psychiatrist prove chatbots cause psychosis?

No. His 2023 editorial was a warning based on reasoning, and the 2026 Aarhus record review found cases but cannot show cause.

Is it only people with mental illness who are affected?

Most reported cases involve people with a history or risk factors, but some reports describe people with no known history. Researchers do not yet know how often that happens.

Where can a family start if they are worried?

Start with a doctor or local mental health service. In a crisis, use 988 or 911.

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