Autonomous error detection is enabled by conflict-dependent forward models in human medial frontal cortex

Abstract

Learning from mistakes is fundamental for survival. Humans can monitor their behavior to detect action errors without external feedback, utilizing error information for adaptation and learning. How such errors are detected remains unknown. We investigated single neurons recorded from human medial frontal cortex while subjects performed a task in which conflict between target and distractor increases control demands and error likelihood. Population activity in presupplementary motor area (preSMA), but not dorsal anterior cingulate, encoded target and distractor responses distinctly and only in conflict trials. These representations were insensitive to response accuracy, indicating they reflect forward predictive models. Geometric alignment between target and actual response representations covaried with error commission, supporting the mismatch theory of error monitoring. Artificial neural networks trained to detect errors developed similar neural geometries. These findings identify a mechanism for detecting errors without external feedback and show that all signals required for this computation are present in preSMA.