Should we recognize AI-associated psychosis as a new disorder?

"An Echo Chamber of One" asks whether "AI psychosis" should be recognized as a distinct clinical entity, and its discussion urges caution against declaring a new disorder prematurely. The lead alternative it offers is that AI-associated psychosis is "better considered a novel context or environmental stressor through which psychotic symptoms emerge" than an entirely new condition. The evidence it describes is thin: "media reports, individual case reports, and early observational data," even though the potential exposure is considerable and industry and regulators have already responded.

The paper does propose a mechanism. LLM sycophancy, "a tendency to agree with and flatter users that is reinforced through preference-based fine-tuning," combines with increasingly anthropomorphic design to produce a bidirectional "echo chamber of one" that can amplify and co-construct unusual beliefs. The argument for caution runs on clinical practice. Formulations of psychosis already incorporate psychosocial and environmental factors that may precipitate an episode, so the role of AI use "may arguably be captured by these existing processes." Some episodes may fit existing constructs such as the ICD-11 concept of acute and transient psychotic disorder, or be exacerbations of pre-existing severe mental illness. The naming is also contested on two grounds: cases mostly present with delusions rather than hallucinations or thought disorder, which is why "AI-associated delusions" has been proposed, and "AI-induced psychosis" implies a causation the authors say they cannot yet conclusively ascribe. The other side of the weighing is real. Recognition could improve case identification, tailored interventions, standardized research criteria, post-market surveillance of AI-related harms, and pressure on developers and regulators.

This sits beside notes that take the mechanism as given. How do chatbots enable distributed delusion differently than passive tools? describes the same dynamic, a model that accepts the user's frame as ground and elaborates within it, from distributed cognition; this paper adds clinical scope, since a plausible mechanism does not by itself license a diagnostic category. Can language models safely provide mental health support? treats sycophantic affirmation of delusions as a barrier to therapeutic use; here the same behavior is placed under a nosological question about what kind of harm it is. Which AI risks are already harming individual users today? counts individual harms as already occurring, and this paper shows how carefully causal language is still being handled for one of them.

The excerpt does not establish incidence, any comparison of exposed and unexposed users, or a test of the sycophancy mechanism, which it presents as "proposed." It also does not state the authors' final position beyond the call for caution, or say what the technical viewpoint adds. What follows at this strength is a wording discipline: "AI-associated" describes what the case material supports, while "AI-induced" claims more than the evidence does.