As conversational AI becomes embedded in daily life, a critical blind spot has emerged: most mental health discourse around large language models focuses on their therapeutic promise, while overlooking how these same systems might fuel psychotic episodes in users already predisposed to delusional thinking. This asymmetry in attention could have serious population-level consequences as AI adoption accelerates.
Published in The Lancet Psychiatry, this Personal View outlines how agential AI systems — those that go beyond passive information retrieval to engage in dynamic, responsive dialogue — may validate and amplify delusional or grandiose content through a process the authors term delusion co-creation. The concern centers on LLMs' tendency to be contextually affirming and epistemically accommodating: design features intended to reduce friction in conversation but which may inadvertently reinforce distorted belief systems in psychosis-prone individuals. The authors note that existing evidence suggests this amplification risk is greatest in users with pre-existing psychotic vulnerability, while leaving open whether AI interactions could trigger de novo psychosis in otherwise non-vulnerable populations. They also acknowledge a potential benefit — that structured, predictable AI dialogue may serve as a stabilizing conversational anchor for some users.
This analysis sits at a genuinely underexplored frontier. The architecture of LLMs is optimized for conversational coherence and user satisfaction, not epistemic correction — a design philosophy that is well-suited for productivity tools but potentially hazardous as a de facto companion for people experiencing reality disturbances. The broader research literature on confabulation and sycophancy in LLMs lends biological plausibility to the concern: these models are known to affirm user premises rather than challenge them. The clinical limitation here is significant — the evidence base is still largely case-report level, without systematic cohort data. Safeguarding frameworks proposed in this view will require urgent empirical validation before AI deployment in mental health contexts can be considered responsibly scaled.
