Phenomenology of AI-Generated "Entity Encounter" Narratives

ChatGPT-3.5-generated narratives of mystical, supernatural, or anomalous entity encounters approximate but do not fully match the phenomenology of real-life accounts. The AI descriptions covered each encounter type, mapped to a Rasch hierarchy of anomalous perceptions, showed below-average scores on the Survey of Strange Events, and referenced at least one recognition pattern of Haunted People Syndrome. Inter-rater reliability was fair, and correlations among narratives were low but generally positive. The findings suggest that prototypical depictions based on popular source material can mimic core features of these experiences, yet they lack the full depth of spontaneous reports.

Study at a glance

Abstract

Objective: We used the ChatGPT-3.5 artificial intelligence (AI)-based language program to compare twelve types of mystical, supernatural, or otherwise anomalous entity encounter narratives constructed from material in the publicly available corpus of information, and compared their details to the phenomenology of spontaneous accounts via the Survey of Strange Events (SSE) and the grounded theory of Haunted People Syndrome (HP-S). Methods: Structured content analysis by two independent and masked raters explored whether the composite AI-narratives would: (a) cover each encounter type, (b) map to the SSE’s Rasch hierarchy of anomalous perceptions, (c) show an average SSE score, and (d) reference the five recognition patterns of HP-S. Results: We found moderate evidence of a core encounter phenomenon underlying the AI-narratives. Every encounter type was represented by an AI-generated description that readily mapped to the SSE, albeit their contents showed only fair believability and low but generally positive correlations with each other. The narratives also corresponded to below-average SSE scores and referenced at least one HP-S recognition pattern. Conclusions: Prototypical depictions of entity encounter experiences based on popular source material certainly approximate, yet not fully match, the phenomenology of their real-life counterparts. We discuss the implications of these outcomes for future studies.