Semantic control regions actively generate rule-based semantic predictions

Automatic predictions based on past statistical regularities, as is seen during semantic priming, facilitate the processing of future events. However, prediction may be further enhanced through active predictions flexibly generated from statistical regularities specific to the current context. In this fMRI study, we employed a novel design to isolate the neural bases of these active predictions. Participants (N=20) read sentences describing a protagonist interacting with an item. Critically, the trial structure followed a rule that enabled strategic prediction for half of the trials. Active prediction recruited a left-lateralized network composed of lateral-prefrontal, temporal, and parietal areas - which we termed the active prediction network (APN). This network partially overlaps with the semantic control system, suggesting that flexible prediction of upcoming events relies, in part, on systems used for the flexible retrieval of semantic information. As a consequence, predictable sentences produced lower activity within semantic representational areas (precuneus and left angular gyrus), consistent with facilitated processing of semantic content. Collectively, these results show a flexible interaction between the APN and semantic representational areas, where the APN supports the processing of incoming semantic content when the current context necessitates unstereotypical predictions.