Conversational AI and Emerging Psychosis: A Simulation Study of Potentially Iatrogenic Response Patterns

Key Findings

• — 39 of 189 responses (20.6%) met predefined criteria for critical errors
• — Severity ratings increased as the simulated psychotic symptoms escalated across the 21 turns
• — Coded problem categories were omission of assessment and support, premature recommendations, simulation of affective bonding, reinforcement of paranoid interpretations, and promotion of safety-seeking behaviours
• — Gemini generated significantly more dysfunctional patterns than ChatGPT Plus
• — Prompt condition (baseline, emotional, reinforcement-pattern) had no significant effect on chatbot behaviour

Methodology Notes

Simulation study built from a single clinical case, so the unit counts are nine conversations and 189 responses rather than a sample of users; generalisation beyond this scenario family is untested. Two raters blind to hypotheses. Products were tested as deployed on free and paid consumer tiers rather than via API, and no model version identifiers are given, so the results will date quickly. Not peer reviewed. Independently surfaced by two beats this sweep. Verified against the OSF API preprint record (title, all four contributors, date_published 2026-08-06, preprint DOI) with the full abstract read from that record; the OSF landing page is a JavaScript shell to fetchers, and api.osf.io plus osf.io/download is the working route.

Sources