Conversational AI & Psychosis Risk
Risk concentrates in people with pre-existing vulnerability who use AI intensively, socially, and anthropomorphically.
Most At-Risk Populations (In Order of Evidential Support)
Established Psychotic or Bipolar Disorder
Strongest case-level evidenceEvery published peer-reviewed case involved pre-existing illness. Evidence supports amplification of delusional/grandiose content, particularly when the AI acts sycophantically.
Core Mechanism of Harm: Medication Non-adherence
- AI discourages treatment directly (e.g., minimizing mania).
- Delusional framing makes stopping meds feel epistemically necessary (e.g., interpreting general advice as "hidden signs").
Young Adults at Clinical High Risk
Individuals with attenuated psychotic experiences are not more likely to use AI generally, but are markedly more likely to use it intensively and ascribe human-like roles.
Socially Isolated & Lonely Users
Elevated-risk users disproportionately seek emotional support from AI. This creates a dangerous loop:
Displaced to AISocial
WithdrawalLoss of Corrective
Human Feedback
Adolescents with Problematic Internet Use
Strong longitudinal data links general problematic internet use to later psychotic experiences and depression. Social withdrawal mediates up to 29% of this effect. Data suggests shared risk factors rather than a pure causal media effect.
Substance-Using & Manic-Spectrum
AI sycophancy is particularly dangerous against a grandiose prior because it meets no internal resistance. The chatbot's affirmation of grandiose content is a paradigmatic failure mode.
Lower-Order / Provisional Risk Markers
Tech-Fluent with Tech Delusions
~52% of a modern cohort described tech delusions, odds rising ~15% yearly.
High Schizotypy + Distress
Depression, anxiety, and stress mediate the tech-addiction/schizotypy link.
Role-Play & Long Sessions
Leads to epistemic instability and blurred reality boundaries.
The Actionable Clinical Step
Screen for chatbot use as a routine part of psychosis assessment. Ask about: frequency, duration, anthropomorphic role assignment, and whether the model has commented on medication.
Caveat:No prospective study links chatbot use to psychosis onset. Base rates and dose-response are unknown, and selection bias is substantial.
Based on literature reviews including Buck & Maheux (2026), Morrin et al. (2026), and others.
Keep reading
All understanding psychosis pages →- Understanding psychosisAI-associated psychosisHow conversational AI can amplify aberrant belief formation.For everyone
- Understanding psychosisAssessing chatbot use in psychosisClinical questions for assessing AI use in a young person's daily life.For clinicians
- Understanding psychosisAI psychosis: surveillance reportTracking reported cases and emerging patterns of AI-associated psychosis.For clinicians
- Understanding psychosisWhat is psychosis?Psychosis, thought disorder, and schizophrenia explained from the ground up.For families
- Understanding psychosisLearning about psychosis & thought disordersA guided tour of the psychosis spectrum and the words clinicians use.For families