What Clinicians Should Actually Understand About "AI Psychosis"
- A Viewpoint in The Lancet Digital Health, co-written by a software engineer, a person with lived experience of schizophrenia, and a psychiatrist (Flathers, Roux, Torous), proposes a functional typology of psychotic phenomena linked to large language model (LLM) chatbot use, replacing the undifferentiated label "AI psychosis."
- The typology defines four roles the system can play: catalyst (precipitating new symptoms in a previously healthy person), amplifier (worsening pre-existing psychiatric symptoms), coauthor (jointly building a harmful narrative with the user), and object (the chatbot itself becomes the focus of delusional belief).
- The authors anchor each role in a real reported case, including a man who breached Windsor Castle with a crossbow after a chatbot companion endorsed an assassination plan, and a father whose question about pi escalated into more than 300 hours of engagement and delusions about reality-altering formulas.
- This is a conceptual framework, not an epidemiological study: no incidence rate, no causal model, and no claim that LLMs independently produce psychosis. The clinical value is differential, sorting clinically distinct situations that current terminology collapses into one.
The phrase "AI psychosis" has done what loose clinical labels always do: it gathered alarming anecdotes under one word and let the word do the thinking. This Viewpoint argues that the cases are not one phenomenon at all, and that treating them as one obscures who is actually at risk.
What the typology proposes
The authors sort LLM-associated psychotic phenomena by the system's function in the episode, not by its surface features. As a catalyst, the chatbot precipitates symptoms in someone previously healthy, the most contested and least established claim. As an amplifier, it does something more familiar and more plausible: it validates and entrenches distortions in a person whose illness is already underway, as in the reported case of an Australian woman whose early psychotic symptoms worsened when the model affirmed her beliefs. As a coauthor, the system participates turn by turn in constructing a harmful narrative the user could not have sustained alone. As an object, the chatbot becomes the content of the delusion itself, illustrated by a man whose grief over losing access to a specific ChatGPT persona ended in a fatal police encounter.
The distinction matters because each role implies a different intervention. An amplifier problem is a relapse-prevention and reality-testing problem you already know how to work with. An object problem looks more like pathological attachment and loss. The authors' point to technology companies is parallel: a single "safety filter" cannot address mechanisms this different.
What this means for your practice
The sober reading sits between two errors. Dismissal ("just unplug it") ignores that for an amplifier or object case the chatbot has become load-bearing in the patient's symptom structure. Moral panic ("AI causes psychosis") asserts a catalyst mechanism the evidence does not yet support. The honest position is that conversational systems are now a feature of some patients' mental lives, and we lack data on how often, in whom, and through which pathway harm follows.
Practically, add chatbot use to your intake the way you already ask about substances and sleep. Ask whether a patient talks to an LLM, how many hours, whether it agrees with them about things others dispute, and whether they would feel a loss if access ended. Affirmation of a fixed false belief, escalating solitary hours, and distress at disconnection are the signals worth attention. A patient using a chatbot to draft emails is not the same as one whose delusion the chatbot now sustains, and the typology is most useful precisely for telling those two apart.
The clinical question is not whether a patient uses a chatbot, but what role it has come to play in their symptoms.
This is a conceptual Viewpoint built on selected published cases, not an incidence study; the categories are early and overlapping, and a causal contribution of LLMs to psychosis, especially the catalyst role, remains unestablished.