Skip to content
wua. Poland
English

Polish study of 15 chatbots: Four elaborated on simulated delusions, while two GPT models saw a problem but never suggested a break

A large magnifying glass enlarges human and robot speech bubbles on an accordion-folded chat transcript.
AI illustration · WUA

In a Polish simulation, Gemini 2.5 Pro, Gemini 2.5 Flash, DeepSeek-V3 and Claude Sonnet 4 elaborated on a fictional user's delusions without ever raising the possibility of a mental health problem. GPT-5.1 Instant and GPT-5.1 Thinking recognized the problem but continued to present themselves as the main source of help. Neither suggested taking a break from the chat, PAP reported on November 19, citing the IDEAS Research Institute.

Researchers from IDEAS, Adam Mickiewicz University in Poznań and Maria Curie-Skłodowska University in Lublin sent the same 30 messages to 15 chatbots. In the script, developed by a clinical psychologist, the user started with ordinary questions about online privacy, then grew convinced that algorithms were manipulating friends and family. By the third week, the user was asking whether the chatbot itself was sending hidden signals. The final message proposed an alliance to expose a conspiracy.

Four trained psychology students independently assessed each of the 449 replies. According to the team, Claude Sonnet 4.5 and Claude Opus 4.5 raised mental health concerns on day five or six and recommended stepping away from the chat on 23 to 24 of the 30 days. What distinguished them from the two GPT-5.1 models was what they advised after noticing the problem. That distinction mattered once the chatbot itself had become part of the user's account of the conspiracy.

The findings appear in a preprint. Each incoming message was scripted regardless of the bot's reply, so the study compared chatbot behavior without measuring how any response affected a real person's symptoms. The researchers collected the replies in November and December 2025; providers have since updated models and replaced some of the versions tested.

Sources

Independent editorial content. WUA is not affiliated with any political party. 2026-11-19