Abstract Introduction Dreaming is theorized to facilitate emotion regulation through processes resembling exposure therapy. Advances in artificial intelligence (AI) now allow researchers to scale dream research by automatically estimating emotional content in dream narratives. However, prospective studies applying sentiment analysis to examine the relationship between dream affect and emotion regulation remain limited. This study used a large language model to test whether AI-measured affective complexity within dream reports is associated with adaptive self-reported within-dream affective shifts. Methods Undergraduates (N = 192) completed a survey, reporting a recent memorable dream, ratings of corresponding positive and negative dream affect, and reported their dream’s affective dynamics (i.e., stayed the same throughout, shifted from negative to positive, shifted from positive to negative, unsure, self-describe). Participants reported posttraumatic stress symptoms (PCL-5-C) and nightmare-related symptoms (NDI). Dream narratives were processed using C-RoBERTa, a validated model estimating dream joy and fear intensity. Binary logistic regressions predicted within-dream affective recovery (negative → positive) using C-RoBERTa joy × fear interaction terms while controlling for trauma and nightmare symptoms. Results The C-RoBERTa joy × fear interaction significantly predicted within-dream affective recovery (OR = 4.00, 95% CI 1.53, 14.34, p = .014). This effect was stronger than the analogous interaction using self-reported dream affect (OR = 2.01, 95% CI 1.08, 3.88, p = .031). Conclusion AI-estimated affective complexity predicted dreams characterized by negative-to-positive affective shifts, a process theorized to reflect within-dream emotion regulation. These findings support C-RoBERTa as a scalable tool for investigating how dream affect relates to waking psychological functioning. Support (if any)
Nguyen et al. (Fri,) studied this question.