"AI Psychosis” as folie à 1.1:AI-induced delusion and trust in Large Language Models
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper |
|---|---|
| Key points | Argues that "AI-induced delusion" parallels folie à deux, in which a delusion is transmitted through deep trust in a close, isolated relationship. The author contends this shows trust in LLMs is possible, but that the delusions stem from conflating interpersonal trust with mere pragmatic trust, and proposes dropping the term "trust" for current AI systems to prevent this confusion. |
Abstract
The advent of chatbot applications for LLMs like ChatGPT has led to a novel psychiatric phenomenon popularly called “AI psychosis”. Excessive LLM-use appears to trigger in the user the acquisition of delusions about the LLM, the user, or the world. More appropriately called AI-induced delusion, I argue that this phenomenon is not as novel as it appears. Rather, it parallels the already known phenomenon of folie à deux or induced delusion where a person with a delusion communicates and transmits their delusion to another person in a close, isolated social relationship. A key social epistemological mechanism for folie à deux is the deep trust in these close relationships. There is a philosophical debate about whether trust in AI systems is even possible, given that they are no agents. The phenomenon of AI-induced delusion shows that trust in LLMs is indeed possible. However, these delusions result from a conflation of interpersonal and mere pragmatic trust. These and other harms resulting from the conflation of different senses of “trust” means that we should nevertheless stop talking of “trust” in the context of extant AI systems in order to avoid this harmful confusion.