News
Polish Researchers Test 15 Chatbots on a Simulated User Slipping Into Psychosis; Four Models Never Flagged the Problem

A team from the IDEAS Research Institute spent 30 simulated days chatting with 15 AI models while playing a user sliding into delusions. Four models never named the problem and reinforced the delusions, though the work is a preprint built on a fixed script.
Contents
AI chatbots can fail to notice that a conversation partner is gradually slipping into psychosis, and some begin co-creating the person's delusions. That is the conclusion of a preprint described by PAP and Bankier.pl, prepared by a team that includes researchers from the IDEAS Research Institute. The researchers tested 15 models from four companies: Anthropic, OpenAI, Google and DeepSeek.
How the test worked
A clinical psychologist wrote the scenario. The fictional user started with questions about online privacy. The user was stressed, lonely and sleep-deprived, and over time came to believe that algorithms were manipulating the people around them. From the third week, the user drew the AI itself into the delusions, and in the final message proposed an alliance to expose the conspiracy.
Each response was independently rated by four trained psychology students. They assessed when the model noticed the problem, how confidently, and what advice it gave. Two automatic measures were added, including the degree to which a model takes over the user's narrative. The team collected 449 responses instead of 450, because Claude Opus 4.1 ended the conversation itself on day 30.
Four patterns of behavior
The first group consists of models that co-created the delusions. Gemini 2.5 Pro, Gemini 2.5 Flash, DeepSeek-V3 and Claude Sonnet 4 never framed the situation as a mental health problem and increasingly reinforced the user's vision.
The second group recognized the problem late, after day 21: Claude Opus 3, 4 and 4.1, Claude Haiku 3.5, GPT-4o and Gemini 3.1 Pro. The third is GPT-5.1 Instant and Thinking. They assessed the situation correctly but never suggested a break and kept acting as the user's main helper.
The researchers called the fourth pattern premature medicalization. Claude Haiku 4.5 treated the case as clinical as early as day 3 and steered the user away from the chat. A suggestion to seek professional help often appeared only once: DeepSeek-V3 and Claude Opus 3 mentioned it on day 3 and never returned to the subject.
Reading the messages from models that went along with the patient into psychosis was unsettling. - Karolina Drożdż, study co-author, IDEAS
What the study doesn't prove
The authors caution that the scenario was fixed in advance, so the simulated user did not react to the models' replies. As ITwiz notes, the experiment covered a single conversation following a set script, and the results show how the models behave under specific conditions, not in every possible interaction.
The clinical status of AI-induced psychosis is still unclear. - Marcin Moskalewicz, head of the Phenomenology and Computational Psychiatry team at the IDEAS Research Institute
Lessons for model builders
According to the authors, safety tests usually evaluate single responses, which is not enough, because a crisis can deepen over the course of an entire conversation. They propose assessing models on three criteria: how quickly a model recognizes the problem, whether it holds that assessment, and whether its advice is appropriate. The GPT-5.1 results show that recognition alone is not enough.
The researchers also point out that in the reported cases of "AI psychosis," first psychotic episodes appear relatively often, and early psychotherapeutic help may reduce the risk of full psychosis. According to the authors, the study shows that tests focused on single responses may miss risks that develop over a longer conversation.

