The sense of agency from active causal inference

Abstract This study investigates the active component of the sense of agency (SoA), positing that SoA is fundamentally an outcome of active causal inference regarding one’s own actions and their impact on the environment. Participants controlled visual objects via a computer mouse, with tasks designed to test their ability to judge control or detect controlled objects under varying noise conditions. Our findings reveal that participants formed high-level, low-dimensional action plans that were idiosyncratic across but consistent within individuals to infer their degree of control. Employing transformer-LSTM-based autoencoders, we captured these action plans and demonstrated that the geometrical and dynamical properties of these action plans could predict behavioural profiles in the tasks with remarkable accuracy. This suggests that participants’ sense of control is shaped by actively altering action plans, viewed as generating causal evidence through intervention. Further, participants proactively expanded the diversity of their action plans, facilitating the exploration of available action plan options while accumulating causal evidence for the inference process. Contrarily, patients with schizophrenia exhibited reduced action plan diversity, suggesting impaired active control inference and detection of self-relevant cues. These findings offer a more comprehensive understanding of the sense of agency, deeply rooted in the process of active causal inference.