AI-Generated Psychiatric Conversations on Hallucinations: Clinical Relevance, Opportunities, and Challenges

This proof-of-concept corpus study examined the technical feasibility of generating and clinically appraising synthetic psychiatric dialogues about hallucinations. A corpus of 200 dialogues was generated with GPT-4o-mini using synthetic personas, structured prompt attributes, and knowledge graph representations. According to the primary clinician’s annotations, hallucinations were rated as present in n=94 dialogues, absent in n=91, and No/Unclear in n=15. Among the hallucination-positive dialogues, auditory phenomena were recorded in n=57, tactile phenomena in n=16, olfactory phenomena in n=14, and visual phenomena in n=13; modalities were not mutually exclusive. The corpus had two clinician annotation layers. The primary psychiatrist provided the annotations used in the descriptive corpus analysis, while a second psychiatrist independently rated the same 200 dialogues in predefined domains for the interrater agreement analyses. Four-category realism showed 75.5% exact agreement (Cohen’s κ=0.549), and three-category hallucination presence showed 75.0% exact agreement (κ=0.552). Exploratory within-corpus comparisons examined associations of the primary hallucination-presence rating with realism and exact symptom labels. The resulting corpus is presented as a structured resource for method development and future educational or computational evaluation, not as a validated clinical or decision-support instrument.