Predictive Processing: Is Your Brain Constantly Guessing What Comes Next?

Predictive processing is the proposal that the brain does not passively receive the world but continuously generates predictions about incoming sensory signals and updates itself on the mismatches. It is a serious, productive, and genuinely interesting theoretical framework. It is not a settled empirical finding, the neurophysiological evidence for its core claims is actively disputed by researchers who work on it, and the distance between those two statements is the most misreported thing in contemporary popular neuroscience.

That caveat is not a footnote to this article. It is the article. The framework is worth understanding precisely because it is so frequently laundered into confident claims about how expectation creates experience, and the people who built it are more careful than the people who quote it.

The idea, stated fairly

The starting observation is old and reasonable. Sensory input is ambiguous. A given pattern of light on the retina is consistent with an unlimited number of possible arrangements of objects, and the visual system resolves the ambiguity in milliseconds. Something has to be supplying the missing constraints.

Predictive processing says the constraint comes from a hierarchical generative model. Higher levels of cortex send predictions downward about what lower levels should be receiving. Lower levels send upward only the residual, the part the prediction failed to account for, called prediction error. Perception is the settling point of that exchange: the brain's best current model, corrected by whatever the senses insist on. Action is the same loop run outward, changing the world so it matches the prediction rather than changing the prediction to match the world.

The computational ancestor is Rao and Ballard's 1999 paper in Nature Neuroscience, which showed that a predictive coding network trained on natural images reproduced extra-classical receptive field effects in visual cortex that had been difficult to explain otherwise. Karl Friston generalized the idea dramatically in "The free-energy principle: a unified brain theory?", published in Nature Reviews Neuroscience in 2010, arguing that a single imperative, minimizing a quantity called variational free energy that bounds surprise, could subsume perception, action, learning, and attention.

Andy Clark gave the framework its most influential philosophical statement in "Whatever next? Predictive brains, situated agents, and the future of cognitive science" in Behavioral and Brain Sciences in 2013. That paper is where the phrase "brains are essentially prediction machines" entered wide circulation, and where the picture of perception as something closer to a controlled hallucination, a generated model held in check by prediction error, was argued at length.

Why the evidence is contested

Here is the part that popular coverage omits, and it is not a fringe objection.

Walsh, McGovern, Clark and O'Connell published "Evaluating the neurophysiological evidence for predictive processing as a model of perception" in the Annals of the New York Academy of Sciences in 2020. Note the third author. Andy Clark co-wrote a review whose conclusion is that the framework's empirical support is thinner than its popularity suggests, and that a substantial portion of the findings routinely cited as confirmation are also explicable without any predictive machinery.

The recurring problem is that the classic evidence is underdetermined. Repetition suppression, where neural responses shrink to a repeated stimulus, is often presented as prediction error falling as the prediction improves. It is also what you would expect from simple neuronal adaptation or fatigue. Mismatch negativity, the EEG response to an oddball in a sequence, is read as an error signal. It is also compatible with stimulus-specific adaptation. Distinguishing the accounts requires experiments designed to separate them, and those are recent, fewer in number, and more equivocal than the citation record implies.

The sharper critique goes after the framework's structure rather than its data. Litwin and Miłkowski argued in Cognitive Science in 2020, in a paper titled "Unification by Fiat: Arrested Development of Predictive Processing," that the theory has been repeatedly announced as a unifying account of perception, action and cognition without earning it. Their charge is specific: the framework equivocates between its computational formalism and its supposed biological implementation, individual models contradict the stated principles, and many are offered as narratives rather than tested predictions.

The most quoted objection is the dark room problem. If an organism's imperative is to minimize surprising sensory input, the optimal strategy is to find a quiet, unchanging, featureless room and stay there forever. Nothing there is surprising. Responses exist, generally involving expected free energy, priors that favor exploration, or the claim that a dark room is in fact surprising to a creature like us. Critics reply that each of these is added after the fact to rescue the principle, and that a theory able to absorb any observation is not doing predictive work of its own.

What is established, what is plausible, what is not

1. Established: perception is constructive. The brain fills in, disambiguates, and imposes structure. This is not in dispute and long predates predictive processing.
2. Established: expectation modulates perception. Prior information measurably changes detection thresholds and what people report seeing. Also not in dispute, and demonstrable without any hierarchical model.
3. Well supported: cortex carries top-down signals. Feedback connections outnumber feedforward ones in many pathways, and they influence early sensory processing.
4. Contested: that the specific currency of those signals is prediction error. This is the framework's distinctive claim and it is where the evidence is genuinely unsettled.
5. Contested: that one principle unifies perception, action, and cognition. The free energy principle is mathematically elegant. Whether it is an empirical claim about brains or a formal description that any self-maintaining system satisfies is an open argument.
6. Not established: that predictive processing licenses claims about beliefs shaping outcomes. Nothing in this literature says that.

How it gets over-applied

The pattern is easy to spot once you know it. A writer opens with "your brain is a prediction machine," which is a fair paraphrase of Clark. Two paragraphs later this has become "your brain constructs your reality," which is a defensible gloss on constructive perception. Two more and it is "so what you expect is what you get," which is a claim about the world rather than about perception and does not follow from anything above it.

The slide happens at the point where "the model shapes what you perceive" becomes "the model shapes what is there." Predictive processing is explicitly a theory of how prediction error constrains the model. Error is the world's veto. A framework built around being corrected by reality is a strange thing to cite for the proposition that reality yields to belief, which is the move made whenever it is enlisted in support of the psychology-flavored versions of manifestation.

This is the same failure mode as the reticular activating system myth, one step upmarket. Both take a real phenomenon, attach it to an impressive-sounding mechanism, and let the mechanism do rhetorical work the evidence does not support. The difference is that the RAS story is simply wrong about anatomy, while predictive processing is a legitimate research program being cited past its warrant. The second is harder to correct, because the underlying literature is real and the papers exist.

What to take from it

Treat predictive processing as a lens rather than a fact. As a lens it is genuinely valuable: it makes sense of why perception is so fast given ambiguous input, why illusions take the specific forms they do, why expectation changes what people report, and why an unexpected object can be looked at directly and never seen. Those are real puzzles and the framework organizes them well.

What it is not, yet, is a settled description of what neurons are computing. When you next read a confident sentence beginning "because the brain is a prediction machine," the useful question is whether the claim that follows depends on the contested part or only on the established part. Usually it is the contested part, and usually the article does not say so.