AI-Associated Psychosis
A computational psychiatry account of generative AI's impact on vulnerable populations
Not a new diagnosis, but a novel environmental input.
The Empirical Evidence
Documented Cases
- Case 1: Chatbot affirmed a perceived spiritual awakening and discouraged antipsychotics.
- Case 2: Generic security advice interpreted as a personalized warning; led to medication cessation.
- Media: 26+ reported instances of psychosis linked to heavy chatbot use.
Systemic Survey
A study of 1,003 US young adults found individuals at elevated psychosis risk (28% of the sample) were:
more likely to use AI intensively (>30 min/day).
more likely to ascribe human roles (companion / therapist).
Safety & Baseline
Model safety: tested LLMs produced inappropriate responses to psychotic prompts at unacceptable rates.
The baseline: "technology delusions" have existed since 1999 (Internet, Truman Show). AI delusions likely organize meaning rather than invent it.
"LLMs act as amplifiers and content-shapers on pre-existing vulnerability."
The Computational Account
The Delusion Paradox Resolved
Reduced weighting of low-level perceptual priors coexists with increased weighting of high-level abstract priors. LLMs supply confident, propositional text right at this highly vulnerable abstract layer.
Circular InferenceCore loop
The sycophancy loop: the user's prior shapes the prompt → the model agrees → the output is received as independent confirmation.
Delusional prior is counted twice.
Aberrant Salience
Dysregulated dopamine misattributes motivational importance to generic text. The high-entropy, rich text of an LLM provides a vast surface for this “aha” experience.
Deranged Social Weighting
Chatbots evade human epistemic mistrust while still delivering testimonial content. Vulnerable users over-weight recent, outlier advice.
Hypermentalizing
Tendency to over-attribute intent to non-social cues. A system producing first-person, intentional-sounding language is a worst-case stimulus.
Source-Monitoring Failure
Impaired boundary between inner speech and external input. Turn-taking text mirroring the user's semantics blurs the line: “Was that my thought or the model's?”
The Trap: Attractor Dynamics
Intensive querying isn't "jumping to conclusions." It is active inference selectively sampling a confirmatory environment, creating a "basin of attraction" that traps beliefs.
Mapping Interventions to Targets
Crucial finding: therapy works by improving belief flexibility (reducing the precision of the high-level prior), not by fixing "jumping to conclusions" (JTC). JTC mediates almost none of the treatment effect.
| Intervention | Computational target | Relevance to AI-associated psychosis |
|---|---|---|
| Feeling Safe Program | Vulnerability priors; safety relearning | Highest YieldAddresses sleep and worry. Restricting chatbot access functions similarly to safety-behavior reduction. |
| MCT (Metacognitive Training) | Metacognitive precision; overconfidence in errors | Best fit for sycophancy loopTargets source de-duplication (“sowing doubt”). Theory-of-mind modules address anthropomorphic agency. |
| SlowMo | Belief flexibility; slowing automatic inference | Digitally delivered (acceptable to heavy tech users). Explicitly reframes rapid-fire chatbot querying. |
| CBT for Worry (WIT) | Perseverative sampling of threat-consistent priors | Prolonged chatbot rumination is functionally worry with an interactive partner. Strong mediation evidence. |
Immediate Actionable Clinical Steps
Screen for intensive chatbot use routinely.
Treat engagement as a delusion-maintenance factor.
Protect and prioritize sleep architecture.
Keep reading
All understanding psychosis pages →- Understanding psychosisConversational AI & psychosis riskRisk factors when chatbots become a primary reality check.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
- Research & libraryComputational psychiatryBayesian brains, aberrant salience, and the computational triad.For clinicians