AI Induced Psychosis: A shallow investigation — LessWrong

lesswrong.com · 15,507 words · saved by 7 readers

Epimistemic status: A small project I worked on the side over ten days, which grew out of my gpt-oss-20b red teaming project. I think I succeeded in surfacing interesting model behaviors, but I haven’t spent enough time to make general conclusions about how models act. However, I think this methodological approach is quite reasonable, and I would be excited for others to build on top of this work! There have been numerous media reports of how ChatGPT has been fueling psychosis and delusions among its users. For example, ChatGPT told Eugene Torres that if he “truly, wholly believed — not emotionally, but architecturally — that [he] could fly [after jumping off a 19-story building]? Then yes. [He] would not fall.” There is some academic work documenting this from a psychology perspective: Morris et al. (2025) give an overview of AI-driven psychosis cases found in the media, and Moore et al. (2025) try to measure whether AIs respond appropriately when acting as therapists. Scott Alexander

x AI Induced Psychosis: A shallow investigation — LessWrong LLM-Induced Psychosis AI Curated 2025 Top Fifty: 14 % 386 AI Induced Psychosis: A shallow investigation by Tim Hua 26th Aug 2025 AI Alignment Forum 32 min read 47 386 Ω 89 “This is a Copernican-level shift in perspective for the field of AI safety.” - Gemini 2.5 Pro “What you need right now is not validation, but immediate clinical help .” - Kimi K2 Two Minute Summary There have been numerous media reports of AI-driven psychosis, where AIs validate users’ grandiose delusions and tell users to ignore their friends’ and family’s pushbac

AI Induced Psychosis: A shallow investigation — LessWrong

saved by

related reading