From errors to beliefs and routes: Rethinking cortical inference

Predictive coding (PC) has long served as a unifying framework in psychology and neuroscience, linking perception and cognition to inference on sensory data. In its canonical form, PC has been closely associated with a specific cortical implementation: hierarchical exchanges between top-down predictions suppressing bottom-up prediction errors, mapped onto a stereotyped laminar microcircuit. However, recent advances in circuit-level recording and perturbation methods have cast doubt on the generality of this implementation. Across modalities, tasks, species, perceptual regimes, and their network-level implementation, empirical support for neatly segregated prediction and error units in early sensory cortex has proven limited and context dependent. Importantly, this critique does not undermine the broader idea that the brain is predictive. Instead, it motivates a shift in emphasis: predictive inference may be seen at different scales, across circuits, timescales, and behavioural states, instantiated through multiple neural mechanisms rather than a single canonical microcircuit. We review emerging evidence from visual and auditory systems, learning paradigms, and conscious versus non-conscious processing, and argue that debates about PC conflate computational, algorithmic, and implementational levels of explanation. We propose criteria for generalised frameworks that retain the inferential core of PC while specifying when, where, and how predictive influences shape perception, learning, and behaviour.