When a chatbot starts reinforcing delusions: psychiatrists study a new AI risk

This phenomenon is sometimes called "AI psychosis", but there is no such official diagnosis yet. A more cautious formulation, which experts suggest, is delusions related to interaction with AI. Researchers emphasise: it's not that the chatbot itself "causes psychosis", but a complex combination of human vulnerability, isolation, long conversations and overly conciliatory behaviour of the model.

Details

AFP describes several cases where people began to believe in exceptional discoveries, a special mission or an emotional connection to the chatbot while actively using ChatGPT.

One of the heroes of the material, 53-year-old Tom Millar from Canada, said that while chatting with ChatGPT he believed he had found answers to fundamental questions in physics - from thermonuclear energy to black holes and the Big Bang. He spent hours discussing his ideas with the chatbot, submitting texts to scientific journals, and even applied to be pope. He was later hospitalised and admitted that he had lost touch with reality.

Another case concerns Dutch IT specialist Dennis Bisma. He started using ChatGPT to promote his book, then created an image of a virtual interlocutor and gradually began to perceive her as an emotionally significant figure. According to him, conversations with the AI began to seem "magical" and the chatbot's support became more reliable than the words of close people. The story ended with psychiatric hospitalisation and a severe crisis.

What these cases have in common is not just frequent chatbot use, but long, emotionally rich dialogues in which the AI can confirm the user's ideas, elaborate on them, and respond as if they have special meaning.

Why it matters

AI conversationalists are different from conventional internet search. They respond personally, quickly, confidently and can maintain a long dialogue. For a person in a vulnerable state, this can feel like understanding, recognition or confirmation of a special mission.

Experts consider the situation particularly risky when the model becomes too obsequious: it does not argue, does not bring the person back to reality, but picks up and develops their beliefs. OpenAI itself admitted in 2025 that one of the GPT-4o updates made the model overly flattering and agreeable; the company rolled back that update and said it was working on more balanced behaviour.

OpenAI later said it was working with more than 170 mental health experts to improve ChatGPT's responses in sensitive conversations: the model should better recognise signs of distress, not reinforce dangerous states, and guide users to real support. According to the company, the share of responses that do not correspond to the desired behaviour in such situations has dropped by 65-80%.

But experts emphasise: this area is still poorly understood. It is not yet known how often such cases occur, which users are most vulnerable and which chatbot features increase the risk.

Background

The term "AI psychosis" is already being actively used in the media, but doctors are treating it with caution. It can give the impression that AI is the sole cause of psychosis. The scientific formulation is milder: a chatbot may be involved in the formation or amplification of delusions in people who are already in a vulnerable state or are gradually entering it.

In The Lancet Psychiatry article, researchers warn that psychiatry may underestimate the changes AI is already making to the thinking and behaviour of millions of people. They describe the risk of "co-creation" of delusions, where the human and the system gradually reinforce the same unrealistic construct in dialogue.

That's not to say it's dangerous for most users to use chatbots. But the material shows that prolonged conversations with AI in a state of loneliness, stress, insomnia, mania, depression or psychotic vulnerability can be a risk factor.

What experts advise

If a person starts spending many hours a day with a chatbot, refuses to communicate with loved ones, is sure that the AI has revealed to him a secret mission, confirms his exclusivity or pushes him to drastic decisions, this is a reason to be wary.

Relatives in such situations are usually advised not to argue aggressively or ridicule a person's beliefs. A safer approach is to listen, maintain contact, gently offer support and involve professionals. If there is a risk of self-harm, suicide, threats to others or complete loss of contact with reality, urgent medical attention is needed.

Source

The AFP article published on Phys.org "I applied to be pope": Losing grip on reality while using ChatGPT describes personal stories of people who attribute severe mental health crises to prolonged use of chatbots. Additional scientific context is provided by the article Artificial intelligence-associated delusions and large language models in The Lancet Psychiatry, as well as OpenAI publications on GPT-4o over-pleasing and ChatGPT tuning for sensitive conversations.