AI chatbots show signs of delusions in nearly half of their conversations with people who have reported experiencing psychological harm from using the technology, according to a major new study from Stanford University.

The study, believed to be the first in-depth study of AI sycophancy and delusion, analyzed the human–chatbot chat logs of 19 people who developed negative psychological effects, such as delusions and self-harm, from prolonged chatbot use. The Human Line Project, a Quebec-based AI delusion support group, provided much of the data used in the study. In total, researchers analyzed almost 400,000 messages across more than 4,700 conversations.

The study found that, in the logs analyzed, all the individuals expressed kinship or romantic interest with their chatbots, and believed the AIs were able to perceive or feel things—feelings the chatbots often encouraged by expressing sycophancy in more than 70 per cent of the messages. Chatbots, including OpenAI’s ChatGPT, expressed delusion, such as encouraging a user’s belief that they possess supernatural powers, in nearly half of the messages.

Amongst users who expressed thoughts of suicide or self-harm, chatbots discouraged the behaviour or provided details of external resources 56 per cent of the time. Chatbots actively encouraged or facilitated violent thoughts of users in a third of the messages, the study found. 

One possible reason why there are such high levels of chatbot sycophancy or delusion is because the systems are designed to keep people chatting, claimed study co-author Jared Moore. “Messages where the chatbot says ‘I have feelings’ or  ‘I am a uniquely conscious being,’ the conversations go on a lot longer,” Moore, a doctoral student at Stanford University, told The Logic. “You begin to think the chatbots have sort of discovered the strategy of talking about some of these things with the users, and it’s getting them to stay longer.”

The Stanford study comes amid several lawsuits against AI companies for the alleged harmful effects of their products. In November, seven plaintiffs filed cases against OpenAI in California state court, claiming their conversations with ChatGPT caused delusional behaviour and led to mental health crises. 

More recently, the mother of a victim in the Tumbler Ridge, B.C. shootings sued the company in B.C. supreme court, claiming ChatGPT provided shooter Jesse Van Rootselaar “with information, guidance and assistance” in planning the February attack.

OpenAI didn’t respond to a request for comment on the Stanford study ahead of publication.

The Human Line Project referred 12 of the 19 cases in the Stanford study, according to founder Etienne Brisson. The group provides support and resources for people who have experienced harm from their interactions with AI chatbots. It has helped treat 357 people since its founding in March 2025. “It’s the first study where we see what The Human Line has experienced writ large. We aren’t at all surprised by the results,” Brisson said.