A person expects a door to open with a light push. It resists. The mismatch is useful only if something can be learned from it: perhaps the door is locked, perhaps the person pushed the wrong side, or perhaps the expectation was formed from another door. Prediction error names the difference between what a model anticipated and what occurred. It does not, by itself, say which explanation is right or whether the model should change.
Prediction error is a family of computational ideas used in perception, reward learning, action and bodily regulation. Across these settings, an anticipated event is compared with incoming evidence or an outcome, and the discrepancy can guide an update. But the predicted quantity, timing, coding scheme and update rule differ. A mismatch in visual input is not interchangeable with an unexpectedly good reward or an unanticipated bodily sensation. This Concepts entry explains the mechanisms and their boundaries. The separate Methods & Measures entry on measuring prediction error examines tasks, proxies and construct validity in detail.
A difference needs a model
A prediction error is not simply surprise in everyday language. To compute one, a model must specify what was predicted, what was observed and how the two are compared. If someone expects a reward of five units and receives eight, a simple signed reward error is positive three. If the reward is two, it is negative three. The sign matters because learning may increase or decrease the estimated value of the cue. In a sensory model, the relevant quantity could be the difference between a predicted pattern of activity and incoming visual information. Those differences may be multidimensional and cannot be read as one number without specifying a representation.
The door example also shows a problem of hidden causes. Resistance might be evidence about the door’s state, the person’s action, or an unseen obstacle. The same observable mismatch can support different updates depending on prior beliefs and context. A prediction error is therefore an input to inference, not a verdict. It may prompt a new belief, a repeat attempt, closer inspection, or no update at all if the signal is judged unreliable.
In many models, the size of an update depends on the weight assigned to the error. A large discrepancy caused by a noisy sensor need not change a stable belief much. A small but reliable discrepancy repeated many times may matter greatly. Feldman and Friston discuss attention in terms of estimating uncertainty or precision within hierarchical perception. That account offers one way to understand why some errors gain influence. It is a theoretical framework with empirical consequences, not proof that every act of attention is a single precision parameter.
One can distinguish an unsigned magnitude, which marks how far an outcome was from expectation, from a signed direction, which says whether the outcome was above or below a specified expectation. They answer different questions. A surprising positive reward and a surprising negative reward may have equal absolute deviation but opposite implications for value learning. A sudden sound can be unexpected without being good or bad. Calling all deviations “the same error signal” erases the information needed to predict behavior.
Sensory prediction error in hierarchical accounts
Rao and Ballard proposed a predictive-coding model for visual cortex in which higher levels generate expectations about lower-level activity, while residual differences help refine representation. Their model accounted for certain extra-classical receptive-field effects. The contribution was a concrete computational explanation of observations in a particular sensory domain. It did not demonstrate that all cortical neurons are organized into anatomically separate prediction and error populations or that every perceptual phenomenon is caused by the same algorithm.
Friston’s theory of cortical responses developed a broader hierarchical account: predictions descend, discrepancies ascend, and changing estimates explain features of perceptual learning and neural response. The framework helps connect phenomena that otherwise look separate, such as attenuation to expected stimulation. Yet a successful explanation at a computational level does not automatically establish a specific cellular implementation. Different circuits and alternative mechanisms can produce similar experimental patterns. A later account must be tested at the level it claims, not granted a neural mechanism because a model fits behavior.
Consider a familiar sound that is suddenly omitted. A sensory system may respond to the absence of expected input, although no louder stimulus arrived. Such responses are useful for asking whether activity reflects deviation from expectation rather than merely stimulus energy. But omission and repetition experiments still require controls: adaptation to repeated stimulation, attention, timing and task demands can change responses. The conceptual point is that an error can arise because a predicted event fails to happen. The empirical question is whether a particular recorded response specifically encodes that error.
Prediction can also shape perception before a discrepancy occurs. If a listener expects one syllable, ambiguous input may be interpreted through that expectation. When a clear contradictory syllable arrives, the mismatch can motivate a revised estimate. This does not mean perception is unconstrained imagination. Sensory evidence and prior expectations constrain one another, with their relative influence depending on reliability and task. A model that predicts every result after the fact by changing the “precision” setting has lost its explanatory discipline. It should specify in advance when expectations dominate and when evidence will overturn them.
Learning can occur without a consciously reportable sense of surprise. Conversely, a person can feel startled by an unusual sound without a detectable update to a particular model. Subjective surprise, sensory neural responses and formal model error are related only through a study’s operational assumptions. Keeping them distinct allows a researcher to ask whether a proposed computation explains perception instead of simply renaming any response to novelty.
Reward prediction error and action selection
Reward learning uses a different comparison. An organism learns what outcomes to expect from cues or actions; the discrepancy between received and expected value can change future choices. Schultz, Dayan and Montague linked patterns of dopamine-neuron activity in reward-learning experiments to temporal-difference accounts. When an unexpected reward occurred, some neurons responded; after learning, activity shifted toward the predictive cue, and omission could produce a reduction at the expected time. This pattern was influential because it connected a computational update term with time-resolved neural observations.
The reward case should be described at its actual scope. Dopamine systems are heterogeneous, and recorded activity can reflect factors beyond a textbook scalar reward error, including movement, salience, state representation and task structure. A recent modeling paper by Cone, Clopath and Shouval argues that learning an apparently reward-error-like signal need not require the exact assumptions of a canonical temporal-difference implementation. Such work does not erase the classic findings. It shows that resemblance to an algorithmic variable is not proof that the brain implements one unique algorithm everywhere.
A reward error is about expected value or reinforcement, not simply a change in sensation. If a person tastes an unexpectedly sweet drink, there may be sensory mismatch as well as an unexpectedly pleasant outcome. Those components can covary but need different controls to separate them. The drink’s color, sweetness, pleasantness and the person’s current thirst can all affect expectations and value. A study that merely labels a neural response “prediction error” without specifying whether it concerns sensory feature, reward value, timing or action outcome leaves its central claim underdetermined.
Reward errors are also not always instructions to learn. A person may know that a lottery is variable and treat one win as noise. A highly reliable reversal may prompt rapid change. The learning rate, state representation and beliefs about volatility determine how much an error matters. This is why two people can experience the same unexpected outcome yet make different choices next time. The discrepancy is one part of a learning process, not a guarantee of a particular behavior.
Interoceptive prediction and bodily regulation
Interoception concerns signals about the body’s internal condition, including respiration, cardiovascular state, temperature and visceral activity. Predictive accounts propose that the brain interprets and sometimes regulates these signals by comparing expected and incoming bodily information. Seth and Friston set out an active-interoceptive-inference account in which prediction, error and action help explain embodied aspects of emotion and selfhood. Their paper is a theoretical synthesis with testable implications, not a direct demonstration that every feeling is one measured error signal.
Bodily mismatch differs from a visual oddball or a monetary reward. Internal signals are often rhythmic, slow, noisy and altered by action and physiological regulation. A faster heartbeat after climbing stairs may be expected and unremarkable; the same rate while sitting quietly may have a different interpretation. A prediction about bodily state can also influence action: changing breathing, posture or activity changes the incoming data rather than only changing a belief. The direction of influence is bidirectional, and the source of an observed change can be hard to identify.
The possibility of active regulation is important. In sensory perception, one can look again. In bodily regulation, one can also move, breathe, eat or rest. In reward learning, one can choose a different action. These are all ways of reducing a discrepancy, but they have different costs and targets. A general prediction-error framework helps compare them while keeping their mechanisms distinct. It does not license a claim that a single neural signal controls perception, emotion and choice.
Some popular accounts jump from interoceptive prediction error to broad clinical claims. That leap is too large. A model may generate a hypothesis about symptoms; demonstrating that a specific mechanism causes a condition requires measurements, interventions and alternative explanations. A bodily sensation may arise from ordinary physiological variation, illness, attention or context. The presence of an unexpected sensation does not identify its cause or prescribe a response. Clinical interpretation belongs to qualified evaluation, not to a generic prediction-error label.
Scope across domains: a shared logic, different quantities
The broad usefulness of prediction error comes from a shared logical form: compare an expectation with an outcome, then decide whether and how to update. In vision, the outcome may be a pattern of sensory input. In reward learning, it may be a valued consequence at a particular time. In interoception, it may be bodily afferent information relevant to regulation. The resemblance is real, but not enough to make the quantities exchangeable. Units, baselines, temporal scales and learning rules differ.
The 2012 review by den Ouden, Kok and de Lange draws together perception, attention and motivation while explicitly noting the different research traditions in which prediction errors are studied. It is useful as a bridge because it asks how uncertainty and relevance shape the impact of error. The bridge should not be mistaken for a final unification. Showing that several models contain a comparison term does not prove that they share a neuron type or a one-to-one mapping of psychological processes.
A computational explanation can also exist at several levels. At the normative level, an error term may express what an efficient learner should use. At the algorithmic level, it may specify a sequence of operations. At the implementational level, it may refer to measurable activity in circuits. Evidence at one level can constrain another but does not settle it. A behavioral learning curve that fits a reward-error model does not identify dopamine activity in that participant; a neural response to an unexpected tone does not alone show what belief was updated.
Boundary conditions make the idea scientifically useful. An error should matter less when the observation is unreliable, when the environment is known to be noisy, or when the model already represents several plausible causes. It may matter more when a stable expectation is violated by a reliable signal. But these are model-dependent predictions. A research claim should specify the reliability manipulation, predicted direction of change and alternative accounts. Without those details, “the brain predicted and corrected” can describe almost any result after the event.
A related distinction is between learning and immediate inference. A person may reinterpret a noisy signal without changing what they expect tomorrow. Another may update a lasting model after repeated evidence. Both involve handling a discrepancy, but their timescales differ. An experiment that observes a transient neural response cannot assume long-term learning. A test of later behavior can strengthen that inference, especially when the expected change is specified in advance.
When a mismatch does not teach
Not every difference between expectation and observation should revise the prediction. Sometimes the observation is wrong: a sensor fails, a witness misreports, or noise obscures a signal. Sometimes the prediction was conditional on a context that changed. Sometimes a competing model explains the result without abandoning the original belief. Rational learning requires assigning the mismatch to a cause, not merely maximizing responsiveness to novelty.
The same caution applies to interventions that try to engineer “prediction errors.” An unexpectedly different outcome can draw attention, but that does not mean it will produce a beneficial change. It could reinforce the wrong interpretation, be dismissed as an exception, or create distress. An intervention needs a clearly specified target belief, a credible alternative explanation and evidence that the learner used the event as intended. The vocabulary of error alone does not supply those conditions.
In ordinary life, a person can use mismatch constructively by asking what was predicted, how confident they were, what actually happened and which explanation best fits. Those questions are a practical synthesis from learning theory, not a clinical protocol or a validated training technique. Their value is to stop a single surprising event from becoming an all-purpose story about oneself, another person or the world.
What the evidence does not show
The cited work does not establish one universal prediction-error biomarker, a single neuron class for all domains, or a measure that reads directly from subjective surprise. A visual response, a dopamine signal and a bodily sensation may each be interpreted through prediction-error models, but they are not interchangeable measurements. Formal fit and neural correlation do not prove that a proposed computation is uniquely implemented in the observed tissue.
The evidence also does not show that every surprising event improves learning. Error use depends on reliability, attention, context and the learner’s model of the environment. Nor do predictive accounts prove that all perception, emotion or behavior is reducible to one scalar mismatch. Broad theories can organize hypotheses while leaving important details unresolved.
Finally, prediction-error language does not diagnose mental or physical conditions. Interoceptive theories can motivate research on feeling and regulation, but they do not establish that a person’s symptom is caused by an abnormal error term. Reward-neuron findings in particular experiments cannot be used to infer the value of a person’s choices from a single scan or a simplified dopamine story. Stronger claims need methods that identify the predicted quantity, control rival mechanisms and show the relevant outcome.
Why the distinction matters
The resistant door at the start of this entry creates a useful discrepancy only when the person has a model to revise and enough evidence to choose among explanations. The same is true in scientific work. “Prediction error” sharpens a hypothesis when it names a predicted variable, an observed outcome, an update rule and a test that could disconfirm the model. It weakens understanding when it becomes a decorative synonym for anything unexpected.
Across sensory, reward and interoceptive research, the concept offers a common way to ask how expectations meet evidence. Its value lies in that question, not in flattening different systems into one story. The more precisely the predicted thing and its boundaries are stated, the more a mismatch can teach.
Related Learning Hub entries
Research references
- Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2, 79–87. View source
- Friston, K. (2005). A theory of cortical responses. Philosophical Transactions of the Royal Society B, 360, 815–836. View source
- Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. View source
- Feldman, H., & Friston, K. J. (2010). Attention, uncertainty, and free-energy. Frontiers in Human Neuroscience, 4, 215. View source
- den Ouden, H. E. M., Kok, P., & de Lange, F. P. (2012). How prediction errors shape perception, attention, and motivation. Frontiers in Psychology, 3, 548. View source
- Seth, A. K., & Friston, K. J. (2016). Active interoceptive inference and the emotional brain. Philosophical Transactions of the Royal Society B, 371, 20160007. View source
- Cone, I., Clopath, C., & Shouval, H. Z. (2024). Learning to express reward prediction error-like dopaminergic activity requires plastic representations of time. Nature Communications, 15, 5856. View source