Predictive processing
A view of the brain as a prediction machine: rather than passively taking in sensory signals and building a picture of the world from them, it constantly generates expectations about what those signals should be and corrects them against the errors that arrive. The mathematics is probabilistic, with predictions flowing down a layered hierarchy and errors flowing back up to revise the model. Its draw is breadth — one principle spanning perception, movement, attention, and emotion — and that same breadth is why critics call it hard to disprove.
The framework, developed across the 1990s and 2000s by Karl Friston, Andy Clark, Jakob Hohwy and others, that treats the brain as a hierarchical Bayesian inference engine constantly generating predictions about its sensory inputs and updating those predictions against prediction errors. The frame originated in the perceptual literature (predictive coding in visual cortex; the Helmholtz machine of Hinton and colleagues) and has been extended to motor control, attention, emotion, and — beginning with Seth 2013 and Barrett and Simmons 2015 — to interoception.
Etymology§
The general idea of perception-as-inference traces to Hermann von Helmholtz in the nineteenth century. Predictive coding in its modern computational form is from Rao and Ballard's 1999 Nature Neuroscience paper on visual cortex. Predictive processing as a unifying frame is from Andy Clark's Surfing Uncertainty (Oxford University Press, 2016) and Jakob Hohwy's The Predictive Mind (Oxford University Press, 2013). Free-energy principle is Karl Friston's preferred technical formulation, developed across a long series of papers from 2005 on. The names are not strictly interchangeable but in the interoception literature they are used roughly synonymously.
Predictive processing is the broad theoretical commitment that the brain is in the prediction business — not passively receiving sensory signal and constructing a representation of the world from it, but actively generating expectations about what those signals should be and updating those expectations as the signals arrive. The mathematical machinery is Bayesian: priors meet likelihoods, posterior beliefs are computed, prediction errors propagate up the cortical hierarchy where deeper layers can revise their generative models. The empirical attractor of the frame is that it integrates perception, motor control, attention, and emotion under a single principle.
For interoception specifically, the predictive-processing frame enters in 2013–2015 with Seth's interoceptive inference and Barrett and Simmons's EPIC model — both adapting the framework to body-state monitoring. The two papers share the predictive-processing commitment but differ in detail: Seth emphasises the inference about expected body state as the substrate of feeling; Barrett and Simmons emphasise the visceromotor cortices (anterior insula, anterior cingulate, ventromedial prefrontal) as the prediction-generating layer. Both place the locus of feeling across a wider network than Craig's insula-centric direct- readout model.
Predictive processing is a framework, not a theory, and the distinction matters for what it can and cannot do empirically. It accommodates a large class of phenomena and has been criticised on those grounds — by Mark Bickhard, by some in the embodied-cognition community, by philosophers of science — for being unfalsifiable in its broadest formulations. The narrower technical-Bayesian formulations (Friston's free-energy principle in particular) make stronger commitments that have generated specific computational and empirical tests. Predictive processing works best as a theoretical umbrella for the post-2013 alternative to Craig, with the umbrella covering a heterogeneous set of more specific positions.