Prediction in action: toward an empirical science of active inference

Active inference has emerged as an influential theoretical framework in cognitive neuroscience, offering a unifying account of perception, action, and cognition under the single principle of sensory prediction error minimization. Yet, despite its growing prominence, the framework is often criticized for the difficulty of extracting qualitatively distinct, falsifiable predictions and its lack of clear empirical grounding. This review addresses these concerns by systematically outlining and evaluating key predictions of active inference across four domains that span the action-perception cycle: action planning, decision-making, motor control, and sensorimotor adaptation. For each domain, we contrast active inference with leading alternative accounts, including optimal (feedback) control theories, utility-based theories, reinforcement learning, equilibrium point hypothesis, and motor command theories. We show that active inference uniquely predicts reliance of goal-directed action on optimistically biased prior beliefs, maximization of information gain when options have equal utility, scaling of random exploration with the precision of policy beliefs, context sensitive encoding of sensory predictions in motor cortex, early and non specific sensory attenuation necessary for movement initiation, and precision weighted scaling of motor output. Overall, we find that while existing work is broadly consistent with some of these predictions, empirical support overall remains fragmentary, as most studies were not designed to distinguish between active inference and competing theories. We highlight key areas where findings are promising, while emphasizing the need for theory-driven experiments that can adjudicate between accounts. In this way, active inference can move beyond a formal mathematical framework toward an empirically grounded theory of brain function.