JMIR Publications

ABSTRACT

Background:

Reports of psychotic symptoms emerging or worsening in the context of sustained interaction with artificial intelligence (AI) chatbots have prompted the rapid adoption of the term “AI psychosis” across clinical practice, media discourse, legislation, and litigation. Despite this activity, no published work has systematically examined how the concept is defined, whether competing definitions converge, or whether existing framings are sufficient to support clinical screening, research classification, or regulatory action.

Objective:

This study aimed to map the definitional landscape of AI psychosis across disciplines and time, identify core characteristics and areas of consensus and divergence, and evaluate whether existing conceptualizations provide operational criteria sufficient for clinical or research applications.

Methods:

Rodgers’ evolutionary concept analysis method was employed to examine how concepts develop and change across disciplinary and temporal contexts. A proportionate search across PubMed, PsycINFO, and Web of Science yielded 306 unique entries, supplemented by citation chaining. Purposive sampling across six disciplinary strata (clinical psychiatry, digital mental health, AI safety, media studies, phenomenological psychopathology, and public health) identified 55 to 65 relevant papers (the range reflects boundary cases where relevance to the definitional question was ambiguous), with 14 anchor texts analyzed in full. Data were extracted for all Rodgers framework elements: temporal evolution, disciplinary variation, core characteristics, antecedents, consequences, synonymous constructs, related concepts, and paradigmatic instances.

Results:

The concept evolved through four phases in less than three years: pre-concept substrate (before 2023), hypothesis formation (2023–mid 2025), clinical recognition and naming (mid–late 2025), and mechanistic modeling with institutional response (late 2025–2026). Six competing framings coexist: AI psychosis as a new diagnostic entity, as a contextual evolution of existing psychosis, as historical continuity, as engineering failure, as an emergent property of human-AI interaction, and as a spectrum of causal involvement. Five core characteristics were identified: delusional content, incorporation of AI, bidirectional reinforcement loops, anthropomorphic misattribution, epistemic destabilization, and social substitution with withdrawal. Nine synonymous terms are in concurrent use, each encoding different causal assumptions. Despite this definitional proliferation, no operational case definition, validated screening instrument, dose-response threshold, prevalence estimate, severity classification, or diagnostic criteria exist for any version of the concept. The concept is already deployed in legislation, litigation, corporate safety, and clinical screening—none of which rest on an agreed-upon operational definition.

Conclusions:

AI psychosis presents a measurement paradox: it is simultaneously over-defined (six competing framings in less than three years) and under-operationalized (no framework sufficient for screening, diagnosis, or research eligibility). This synthesis proposes a working definition that accommodates existing causal taxonomies, centers human–AI interaction as the unit of analysis, and identifies eight measurable dimensions suitable for instrument development. Until the field resolves this paradox, the concept will continue to be used in policy, practice, and litigation without the definitional foundation those applications require. Clinical Trial: NA