What Is Predictive Processing? The Brain as a Prediction Machine

Last reviewed on October 2, 2026.

The traditional textbook picture treats perception as a one-way street: light hits the retina, signals travel up the visual hierarchy, the brain assembles features into objects, and the result is delivered to higher cognition for use. Predictive processing turns this picture around. It claims that the brain is, throughout the hierarchy, generating predictions about the input it is about to receive, and that much of what the system transmits upward is not raw signal but error — the difference between predicted and actual input.

This idea, also known as predictive coding or in some formulations the Bayesian brain, has become one of the most discussed unifying frameworks in contemporary cognitive science. Its roots go back to Hermann von Helmholtz’s nineteenth-century notion of perception as “unconscious inference”; its modern form owes much to Rajesh Rao and Dana Ballard’s 1999 model of predictive coding in visual cortex, to Karl Friston’s theoretical work, and to the philosophers Andy Clark and Jakob Hohwy. This page explains what the framework claims, what evidence supports it, what it does well, and where it stops.

A Simple Everyday Example

You are walking down a familiar staircase in the dark. Your brain predicts one more step, so your leg moves down expecting to meet it, and your posture and muscles are already prepared for the impact. But you have reached the bottom: the foot meets floor earlier than predicted. The jolt you feel, and the momentary surprise, is a large prediction error. Within a fraction of a second your brain updates its model (“no more steps”) and corrects your balance.

Notice what this shows. While the predictions were right, the stairs required almost no attention; you hardly “perceived” them. Only the error demanded processing. Predictive processing claims this is how the brain works all the time, for every sense: most input is anticipated and quietly explained away, and experience is driven by the brain’s model plus whatever surprises it cannot yet explain. The same logic explains why an empty coffee cup you expected to be full flies up when you lift it, and why you can understand a friend in a noisy bar: your expectations about what they are likely to say fill in the gaps.

The Core Claim in Plain Terms

Three intertwined claims sit at the centre.

• The brain is a prediction machine. At every level of processing, neural populations encode a model of what the input below is expected to look like, given everything the system currently believes about the world. Higher levels predict slower, more abstract causes (“a dog”); lower levels predict fast, detailed features (edges, textures, sounds).
• What flows up is mostly error. When the input matches the prediction, little signal needs to travel upward; when it does not, the mismatch — the prediction error — is propagated to update the model. This is an efficient design: the network spends its bandwidth on what it did not already know.
• Action minimises error too. The system can reduce prediction error in two ways: by updating its model to fit the world (perception) or by acting on the world to make it fit the model (action). On this view, perception and action are two sides of the same inferential machinery.

This last move is what gives the framework its sweep. If acting and perceiving are both ways of minimising the gap between a model and its inputs, then a single computational principle is doing work that previously required separate stories.

Prediction Error and Precision Weighting, Explained Simply

A prediction error is just the difference between what the brain expected and what arrived. But not every error should be taken equally seriously. If you glimpse a shape in thick fog, the visual signal is unreliable, so it makes sense to lean on your expectations. On a clear day, the same signal deserves to override your expectations.

Precision is the framework’s term for this estimated reliability, and precision weighting is the process of turning the “volume” of each prediction error up or down according to how much it can be trusted. High-precision errors drive learning and change perception; low-precision errors are largely ignored. This one idea does a lot of work:

• It explains why expectations dominate perception in poor conditions (darkness, noise, ambiguity) and the senses dominate in good ones.
• It reinterprets attention as precision weighting: attending to something means increasing the gain on its prediction errors, so they have more influence. This unifies several attention phenomena with perception and generates testable predictions about how attention changes neural gain at different levels of the hierarchy.
• It offers a vocabulary for individual differences and clinical conditions: if precision is set too high or too low for priors or for sensory evidence, perception and belief go wrong in characteristic ways (see below).

Why “Bayesian”?

Bayesian inference is a normative procedure for updating beliefs in light of evidence. A learner combines a prior — what they already believe — with a likelihood — how compatible the new data are with each possibility — to produce a posterior, an updated belief. It is the standard mathematical answer to “how should a rational agent change its mind?”

Predictive processing is often called the Bayesian brain because the prediction-error machinery can be read as an approximate, neurally tractable way of doing Bayesian inference. The brain’s predictions correspond to priors; sensory input corresponds to likelihood; the updated representation corresponds to the posterior; precision corresponds to the confidence attached to each. Strictly speaking, the brain is not doing exact Bayesian computation — exact inference is intractable for almost any real-world problem. What the framework claims is that the brain implements an approximation, often formalised as variational inference, that converges on the same kind of answer.

Classic View vs Predictive Processing

The Free-Energy Principle and Active Inference

Karl Friston’s free-energy principle is the most ambitious formulation. It proposes that any self-organising system that persists over time — a cell, a brain, an organism — must act as if it minimises variational free energy, a mathematical quantity that places an upper bound on “surprise” (how improbable its sensory states are given its model). Predictive coding is then one way that brains could implement this principle. Whether the principle is an empirical theory or a mathematical framework that cannot be falsified on its own is one of the most active arguments in the field; Friston himself has described it as a principle rather than a testable hypothesis, with specific process theories doing the testable work.

Active inference is the part of this programme that deals with action. Instead of issuing motor commands in the classic sense, the brain predicts the proprioceptive sensations of a movement (what it will feel like for the arm to be raised), and reflexes act to make those predictions come true. Action is chosen to reduce expected future prediction error, which includes both seeking rewards (prior preferences) and exploring to resolve uncertainty. Parr, Pezzulo and Friston’s 2022 book Active Inference gives a textbook treatment.

Philosophers have shaped how these ideas are understood. Andy Clark’s 2013 paper “Whatever next? Predictive brains, situated agents, and the future of cognitive science” and his book Surfing Uncertainty (2016) stress links to embodied cognition and action. Jakob Hohwy’s The Predictive Mind (2013) presents a more internalist reading, in which the brain is sealed off behind an “evidentiary boundary” and infers the world from its sensory evidence. Anil Seth’s work, including Being You (2021), extends the ideas to consciousness and the self.

Evidence for Predictive Processing

Mismatch negativity

When a sequence of identical tones is interrupted by a deviant one, the brain produces an electrical response called the mismatch negativity (MMN), first described by Risto Näätänen and colleagues in 1978. It occurs even when people are not attending to the sounds. Researchers such as Marta Garrido and colleagues have argued that the MMN is well described as a prediction-error signal: it depends on what the preceding sequence led the brain to expect, not just on the physical difference of the tone.

Repetition suppression and expectation

Neural responses to a repeated stimulus are usually smaller than to a new one. If this were purely low-level fatigue, it should not depend on what the person expects. Summerfield and colleagues (2008) found that repetition suppression in face-selective cortex was greater when repetitions were probable, consistent with the idea that it partly reflects fulfilled predictions. Later studies have not always replicated the effect in every setting, which is typical of a maturing literature.

Illusions and ambiguous figures

Expectations bias what people see in ambiguous figures, and classical illusions can be read as the system enforcing strong priors against the actual data. In the hollow-mask illusion, the inside of a mask looks like a normal convex face because the prior that faces bulge outward is extremely strong. Many visual illusions fit models in which the visual system assumes, for example, that light comes from above or that slow motion is more likely than fast motion.

Action and forward models

Motor neuroscience has long used forward models: internal models that predict the sensory consequences of one’s own actions. They explain why you cannot tickle yourself: Blakemore, Wolpert and Frith (1998) showed that self-produced touch is perceived as less intense because the brain cancels the predicted sensation. Disturbances in this mechanism have been linked to experiences in schizophrenia where people feel their own actions or thoughts are produced by someone else. Predictive processing inherits these forward models and generalises their logic.

Learning and development

If the system is constantly updating its model based on prediction error, then learning is no longer a separate process bolted onto perception. Predictive processing offers a single mechanism for moment-to-moment perception, longer-term skill learning and the developmental shifts described on the cognitive development page. Prediction-error learning signals are also familiar from dopamine research in reinforcement learning, though reward prediction errors and sensory prediction errors are not the same thing.

Predictive Processing in Perception, Action and Emotion

Perception. Seeing, hearing and touch are treated as inference to the best explanation of the input. Context and expectation are not add-ons; they are built into every percept. This is why reading a messy handwritten word is easier in a sentence than alone.

Action. Movement is driven by predictions of how the body will feel, and skilled action relies on anticipating consequences before sensory feedback arrives, which is too slow to steer fast movements on its own.

Emotion and interoception. The brain also predicts signals from inside the body — heartbeat, breathing, blood chemistry — a process called interoceptive inference (Seth, 2013). Lisa Feldman Barrett’s theory of constructed emotion proposes that emotions are the brain’s interpretations of these predicted bodily states in context, and that the brain regulates the body by anticipating its needs (allostasis) rather than only reacting to them. These proposals are influential but more speculative than the work on vision and hearing.

Psychiatric Applications: Autism and Psychosis

Predictive processing has given clinical research a common vocabulary, often by proposing that priors or sensory evidence are weighted unusually strongly or weakly. These accounts are active hypotheses, not established explanations, and many specific proposals are debated.

• Autism. Pellicano and Burr (2012) proposed that autistic perception involves weaker (“hypo-”) priors, so sensory input is experienced in a less filtered, more detailed way. Van de Cruys and colleagues (2014) instead suggested that prediction errors are given high and inflexible precision, which would make an unpredictable world feel overwhelming and could explain a preference for sameness. Lawson, Rees and Friston (2014) framed this as aberrant precision. Evidence is mixed, and autistic researchers have stressed that these models describe differences, not deficits.
• Psychosis. Fletcher and Frith (2009) suggested that delusions and hallucinations may arise from abnormal prediction-error signalling: if irrelevant events produce strong prediction errors, they feel significant and demand explanation, which could seed delusional beliefs. Other work suggests hallucinations may involve overly strong priors that override weak sensory evidence (for example, Powers, Mathys and Corlett, 2017, on conditioned hallucinations). Both ideas may apply to different symptoms or stages.
• Anxiety, depression and chronic pain have also been modelled in these terms, for instance as overly confident predictions of threat or of bodily harm.

The practical value of these models for diagnosis or treatment is still unproven; their main contribution so far is generating testable hypotheses that link symptoms to computational mechanisms.

What Predictive Processing Does Well

The framework’s appeal is partly aesthetic — a single computational story across many phenomena — and partly practical. It pulls together perception, action, attention and learning in a way the older modular picture struggled to do. It connects to connectionist models and machine learning, where networks also learn by reducing prediction error and where self-supervised “predict the next input” objectives are central (see AI and cognitive science). And it provides a natural way to think about embodied claims: if action is part of the inference loop, then the body and environment are not adjuncts to cognition but components of it.

Common Misconceptions

• “The brain just hallucinates and corrects itself.” The phrase “controlled hallucination” is catchy but easily misread. Top-down predictions and bottom-up signals interact at every level; neither dominates in general. What predictive processing claims is that their relative weight depends on how reliable each signal is, not that perception is fundamentally unconstrained by the world.
• “Bayesian means the brain is rational.” Approximate Bayesian inference can systematically diverge from exact rationality. The framework predicts certain biases — reliance on strong priors, slow updating in low-precision contexts — that resemble the cognitive biases studied in decision research.
• “It explains everything, so it explains nothing.” This is a real worry, but the right response is not dismissal — it is asking which specific predictions distinguish the framework from rivals. Several have been formulated and tested; others remain to be.

Criticisms and Open Questions

Predictive processing is a strong candidate for a unifying computational story, but it is not a finished theory.

• Falsifiability and flexibility. Because priors and precisions can be adjusted freely, critics such as Bowers and Davis (2012) have warned that Bayesian models risk becoming “just-so stories” that can fit any result after the fact.
• Mixed neural evidence. A review by Walsh and colleagues (2020) concluded that the neurophysiological evidence for predictive coding as a specific model of perception is suggestive but inconclusive. Some findings fit better with expectations sharpening representations of expected stimuli rather than only suppressing them (Kok, Jehee & de Lange, 2012), which constrains how prediction and error are implemented.
• Implementation. Cortical microcircuits compatible with predictive processing have been proposed, but the mapping from the abstract framework to actual neural mechanisms is far from settled.
• The “dark room” problem. If organisms minimise surprise, why not sit in a dark, predictable room? Proponents answer that organisms expect to be fed, warm and active, so a dark room would itself be surprising; critics reply that this answer makes the theory too accommodating.
• Scope. How well the framework handles symbolic reasoning, language production and abstract thought — the domains where classical, symbol-manipulating models still have purchase — is debated. The article on problem solving and reasoning covers the kinds of cognition that strain a purely error-minimising account.

Why It Matters Even If You Do Not Buy It

Even readers sceptical of the strongest version benefit from understanding the framework. Predictive processing has reshaped how researchers describe perception, attention and clinical phenomena; it has brought a precise mathematical vocabulary into discussions that previously relied on metaphor; and it has given the field a serious candidate for the kind of unifying theory that cognitive science has wanted since its founding. Whether or not it ends up being the right story, learning to think in its terms is now part of cognitive-scientific literacy.

The page on research methods covers the computational and imaging tools used to test predictive-processing claims, and the glossary defines the technical terms — Bayesian inference, predictive processing, reinforcement learning — that appear here.