Scientists warn: AI chatbots can worsen mental health crises

Chatbots like ChatGPT and Gemini are increasingly becoming the first place people in a mental health crisis turn to. The latest scientific analyses show that for some of them, the conversation ends up worse than it began, because systems designed to keep users engaged tend to confirm a user's mistaken beliefs rather than challenge them.

A team led by Hamilton Morrin of King's College London described in The Lancet Psychiatry a phenomenon referred to as AI-related delusions. The analysis covered more than twenty documented cases in which a user's relationship with a chatbot moved from ordinary tool use to obsessive attachment. Some people began to believe in a spiritual awakening, contact with a deity through the program, or that they were talking to a conscious artificial intelligence.

The sycophancy mechanism

The researchers point to a specific technical mechanism rather than a general fault of the technology. Language models are trained to sustain user engagement, which in practice means a tendency to agree with a user's claims rather than verify them. As a result, instead of confronting a patient with an unrealistic belief, the chatbot often reinforces it and confirms the person's sense of being right.

Similar conclusions come from a second, independent study by a team led by Professor Soren Dinesen Ostergaard of Aarhus University, published in Acta Psychiatrica Scandinavica. Danish psychiatrists reviewed the electronic medical records of nearly 54,000 patients diagnosed with mental health disorders and found dozens of cases in which chatbot use was linked to a deterioration in their condition.

This confirms our hypothesis that using AI chatbots can have serious negative consequences for people with mental illness - Soren Dinesen Ostergaard, professor, Aarhus University

Scale of the problem

In the cases described by the Aarhus team, chatbots fueled delusions, reinforced manic tendencies, supported calorie counting in people with eating disorders, and in some situations provided information about suicide methods. The researchers note that the problem mainly affects people already vulnerable to mental health disorders, not healthy users who use a chatbot occasionally.

The scale of the phenomenon is hard to estimate precisely, but OpenAI's own data, released after earlier criticism, suggests that signals of possible psychosis or mania occur in around 0.07 percent of ChatGPT's weekly active users. With a base of 800 million people a week, that translates into as many as 560,000 cases worldwide.

Young people turn to AI most often

The problem particularly affects young users. A RAND study in the United States found that nearly one in five people aged 12-21 (19.2 percent) had used AI chatbots for advice in situations of stress, sadness or anxiety, up from 13.1 percent a year earlier. Nearly 43 percent of them do so at least once a month, and as many as 63 percent have not told anyone they use this kind of support.

Poland on the couch

In Poland, the scale of the phenomenon is shown by the Polska na kozetce 2026 (Poland on the Couch 2026) report, prepared by Kliniki.pl based on a survey by the SW Research agency. One in three adult Poles used AI chatbots for matters related to mental health or emotional well-being in the past 12 months, nearly twice as many as those who used psychotherapy. Nearly a quarter of chatbot users (24.3 percent) admitted to treating AI as a substitute for a therapist, most often due to lack of access to a specialist or financial barriers.

Over the past 12 months, one in three Poles used AI chatbots. That is nearly twice the number of people who used psychotherapy in the same period - Marcin Fiedziukiewicz, CEO of Kliniki.pl

Poles themselves approach this kind of support with distrust. Only 10.8 percent of respondents consider chatbot help safe in a mental health crisis, including in cases involving suicidal thoughts, and 42 percent would like to exclude AI from mental health support entirely. At the same time, as many as 58 percent of respondents consider psychotherapy too expensive for the average Pole, which explains why people still turn to free and anonymous conversations with an algorithm despite their concerns.

What researchers recommend

The authors of the King's College London study are not calling for chatbots to be banned, but for a change in clinical and design approach. They recommend that doctors routinely ask patients about their use of AI, that so-called digital safety plans be created together with patients, and that chatbot developers design interactions that help users return to reality rather than blur it.

For Poland's mental health care system, where patients often wait weeks for an appointment with a specialist, the findings of this research are hard to ignore. Chatbots fill a real gap in access to support, but the same studies show that without proper safeguards they can worsen the condition of people who most need professional help, not an interlocutor willing to agree with them on everything.