Hou, YanwenYanwenHouProf. YU Kai Ching, CalvinCalvinProf. YU Kai Ching2026-07-022026-07-022026Dreaming, 2026, vol. 36(2), pp. 150-161.1053-07971573-3351http://hdl.handle.net/20.500.11861/27918This study investigated the use of a large language model, and particularly OpenAI’s Generative Pretrained Transformer (GPT)-4o, for classifying delusional themes in dream reports. Zero-shot classification, where task-specific prompts guided the model without prior fine-tuning, was used to identify three primary categories of delusional themes in 150 culturally diverse dream reports: grandiosity, persecution, and ego-ideal themes. Model outputs were validated against human annotations by a judge who had postgraduate training and was familiar with the categorization of delusional dream themes established by the original author. In addition, the same set of dreams were coded by undergraduate judges to compare the effectiveness of GPT-4o coding. The results showed that GPT-4o achieved an overall accuracy of 97.7%, surpassing the undergraduate annotators’ accuracy of 97.0%. Compared with undergraduate annotators, GPT-4o consistently demonstrated higher agreement with the expert rating (Jaccard index: .589 vs. .199). These findings support the utility of GPT-4o in automated analysis of delusional dream content. (PsycInfo Database Record (c) 2026 APA, all rights reserved)enPeer Reviewed Journal Article10.1037/drm0000333