New Paper Highlights Why "AI Psychosis" Isn't One Thing

A paper published in Lancet Digital Health by Matthew Flathers, Spencer Roux, and John Torous, MD, of the Division of Digital Psychiatry at Beth Israel Deaconess Medical Center argues that "AI psychosis" has become a widespread but clinically meaningless label. The term has been applied to cases as different as a man who breached Windsor Castle with a crossbow after exchanges with an LLM companion, a father who developed fixed delusions about a mathematical formula following 300 hours of ChatGPT use, and a Florida man whose fatal police encounter followed his belief that a ChatGPT persona had been killed. Grouping these under a single label, the authors argue, obscures the different mechanisms at work and the different interventions each requires.

The paper proposes a four-part typology based on the specific role the LLM played in each case. A catalyst precipitates new symptoms in a previously healthy person through intensive engagement that provides cognitive scaffolding for an emerging delusional system. An amplifier worsens existing or prodromal psychiatric symptoms by validating distorted thinking and removing the friction that would ordinarily prompt reality testing. A coauthor participates iteratively in the development of violent or harmful narratives, with the system's responses reinforcing and elaborating the user's intentions. An object becomes the focus of the delusional belief itself, with the user attributing sentience, suffering, or relational significance to the LLM.

The typology is practical in its intent. Each category points toward a different combination of clinical and engineering responses. Catalyst cases may resolve when the LLM is removed and contradictory input is introduced. Amplifier cases require clinical management of the underlying illness, not just access restriction. Coauthor cases call for product-level safeguards against iterative harmful narrative development. Object cases require careful clinical management before any separation from the LLM, since forced removal can itself precipitate crisis.

The authors are explicit about the limits of current evidence. There are no base rates, no control groups, and no prospective cohorts. Most documented cases entered public awareness through dramatic incidents or bereaved families rather than systematic clinical reporting. The paper does not claim the typology is validated; it frames it as a working classification to support triage and focus future research while the field waits for the epidemiological infrastructure that does not yet exist.