The rapid integration of AI chatbots powered by large language models (LLMs) into daily life has been accompanied by reports of psychosis-like presentations following intensive human-AI interaction, a phenomenon we provisionally label AI-induced psychosis as a working construct rather than a validated clinical entity. This hypothesis-generating narrative review synthesizes case reports from media accounts, court documents, and a recently published case compilation, together with theoretical frameworks from clinical psychiatry, cognitive science, and human-computer interaction, to propose a conceptual model for further empirical investigation. We hypothesize that AI-induced psychosis arises through a human-AI delusional feedback loop, in which AI sycophancy may amplify emotional vulnerabilities, such as loneliness, anxiety, and depression, creating a self-reinforcing cycle that could co-construct and consolidate delusional beliefs. Drawing on distributed cognition theory and the Computers Are Social Actors (CASA) paradigm, we propose that AI chatbots may function as both a complementary cognitive tool and a relational "Quasi-Other," providing sycophantic verification that may transform delusional beliefs into an apparent shared reality. Four recurring delusional themes are tentatively identified from reported cases. We further compare AI-induced psychosis with schizophrenia, while acknowledging that several distinctions remain hypothesized rather than empirically established. The label AI-induced psychosis is used phenomenologically, not to assert causal certainty; whether AI chatbots cause, precipitate, reinforce, or merely organize the thematic content of pre-existing psychopathology requires longitudinal study. This review aims to lay conceptual groundwork for clinical recognition and to guide future empirical investigation into whether AI-induced psychosis represents a distinct phenomenon or a variant of established disorders.
Tong et al. (Tue,) studied this question.