Mapping content and dynamics in the stream of consciousness through latent brain state analysis

Human thoughts often arise unprompted, forming a stream of consciousness. Despite their ubiquity, it remains unknown how the brain supports unprompted thoughts as an integrated experience with rich content and dynamics. Applying latent state modeling to fMRI data collected while individuals spoke aloud their thoughts, we identified recurring brain states linked to moment-to-moment thought content. A brain state with activation in core default mode network (DMN) regions tracked early emergence of internally oriented thought, while a state involving executive control and dorsal attention networks (DANs) tracked externally oriented thought. Another state involving the medial temporal lobe and DAN engaged later during internal thought verbalization, suggesting roles in elaboration. The latter state was also associated with a large-language-model derived surprisal metric, perplexity, signaling novelty and internal event boundaries. Our findings reveal a multistage architecture of unprompted thoughts, where DMN subsystems interact with attention and control networks to scaffold their fluid progression.