You expect the stairs to be easy.
Halfway up, your breathing is heavier than expected and your heart is working harder. Something in the system has changed before you have a story about it.
One family of theories describes this kind of event as a mismatch between predicted bodily input and incoming bodily input.
The mismatch is often called an interoceptive prediction error.
That phrase has become influential in neuroscience, psychiatry and theories of emotion. It is also easy to misuse. A prediction error is not simply “a surprising feeling in the body,” and it is not a diagnosis. It belongs to a computational framework in which perception is treated as inference under uncertainty.
Definition
In predictive-processing accounts, the nervous system is not understood as passively receiving sensory data and then interpreting it from scratch.
Instead, higher levels of a processing hierarchy generate predictions about likely sensory states. Incoming signals are compared with those predictions. Mismatches can update the model, alter attention, change action or be explained away depending on their estimated reliability.
When the sensory domain is the internal body, the mismatch is described as an interoceptive prediction error.
This may concern expected heartbeat, breathing effort, visceral state, temperature, energy availability or other internal variables, depending on the model and experimental design.
The term therefore names a computational relation, not a discrete sensation.
Prediction is not conscious expectation
A person does not need to think, “My heart rate should now be 72 beats per minute,” for predictive processing to be relevant.
Predictions in these models are often implicit, hierarchical and distributed. They can reflect learned regularities about what bodily state normally accompanies standing, running, eating, resting, danger or illness.
The conscious feeling of surprise may be one consequence of mismatch.
It is not the definition of prediction error itself.
This distinction matters because popular explanations often collapse computational prediction into conscious belief.
The theory is broader than that.
From prediction error to updating
A mismatch can be handled in more than one way.
The system may update its prediction. It may act on the body or environment to make sensory input better fit the predicted state. It may redirect attention. It may change the estimated reliability of the signal.
Active-inference accounts emphasize this reciprocal structure: organisms do not only revise beliefs to match the world; they also act in ways that change sensory states.
If thirst predicts a need for fluid, drinking changes the bodily state. If standing produces unexpected dizziness, sitting down changes cardiovascular demand. If a breath feels unexpectedly difficult, behavior and attention may change before a conscious explanation is settled.
The organism is part of the inference loop.
Precision weighting
Not every mismatch should have equal influence.
A useful computational idea is precision weighting: the system estimates how reliable a prediction or prediction-error signal is and adjusts its influence accordingly.
If incoming sensory evidence is noisy, prior expectations may dominate. If sensory evidence is considered precise, it may drive larger updates.
This makes predictive-processing models more interesting than a simple “prediction versus reality” metaphor.
The question becomes: which signal gets trusted, and how much?
That has obvious relevance to interoception, where internal signals can be noisy, ambiguous and state-dependent.
An example: heartbeat and arousal
Imagine a person entering a meeting where they expect conflict.
Their heart rate rises. That increase may partly match prior expectations of threat. If the bodily change is stronger or earlier than predicted, mismatch may occur. Attention may shift inward. The person may interpret the arousal as evidence that the situation is dangerous.
But the chain is not deterministic.
The same cardiovascular change might be interpreted as readiness, caffeine, heat, social evaluation or physical exertion. External context and learned concepts contribute to meaning.
Prediction error does not tell us which interpretation is correct.
It describes one possible computational ingredient in updating.
Interoceptive inference and emotion
Anil Seth and colleagues developed influential accounts of interoceptive inference, proposing that emotional and bodily self-experience can be understood partly through predictive regulation of internal states.
Later work has elaborated these ideas within active inference, allostasis and computational psychiatry.
Greenwood and Garfinkel's 2025 review places predictive accounts alongside other levels of interoceptive processing rather than treating them as the whole field.
That is the safer position.
Prediction is one theoretical lens for understanding how internal sensory signals and prior models interact.
It does not erase anatomy, physiology, learning or context.
Experimental evidence is growing, not complete
Human experiments increasingly attempt to manipulate or measure interoceptive prediction errors.
Ventura-Bort and Weymar's 2024 study examined whether transcutaneous auricular vagus nerve stimulation modulated neural processing associated with interoceptive prediction error and allostatic regulation. Such work is valuable because it moves the theory toward experimentally testable predictions.
But one study or one neural signal cannot establish that the whole predictive-processing framework is correct.
Researchers still debate how prediction error should be operationalized, which neural signals correspond to it, and how specific computational parameters map onto subjective experience.
A model gains scientific value by making constrained predictions, not by being able to redescribe every outcome after the fact.
The clinical appeal
Predictive models are attractive in psychiatry because they can link bodily symptoms, expectations, attention and learning.
Recent reviews have used interoceptive computational frameworks to interpret anxiety and depression. For example, a 2025 review of anxiety disorders proposed that overly precise threat-related priors and rigid context models may amplify or mis-handle interoceptive mismatch. A 2025 active-inference account of depression linked mood and bodily symptoms to altered long- and short-timescale predictions about internal state.
These are explanatory models.
They should not be presented as settled biomarkers.
Anxiety is not “caused by prediction error” in a single proven sense. Depression is not reducible to one computational parameter. Clinical conditions are heterogeneous and involve biological, psychological and social processes beyond interoceptive inference.
Prediction error is not error in the ordinary sense
The word error can sound like failure.
In computational models, it means discrepancy.
Prediction error is necessary for learning. A system that never registered mismatch could not adapt when circumstances changed.
The relevant question is not how to eliminate prediction error.
It is how mismatch is generated, weighted and resolved.
Too little influence from reliable sensory evidence could make a model rigid. Too much influence from noisy signals could make experience unstable. But even this language is a model-level simplification, not a clinical prescription.
Interoception and exteroception should not be separated too crudely
Interoception is often defined as sensing the internal body and exteroception as sensing the external environment.
A 2024 review by Toussaint, Heinzle and Stephan argues for a computationally informed distinction rather than a purely anatomical one.
That matters because bodily regulation depends on both domains.
The same rapid heart rate means different things while sprinting, lying in bed with fever, standing on stage or sitting quietly after a frightening message. External cues shape predictions about internal state, and bodily state changes how external signals are interpreted.
The inference loop crosses the skin conceptually even when sensory pathways remain distinguishable.
Why “trust the prediction error” is the wrong lesson
Some popular accounts treat mismatch as a hidden truth signal: the body “knows” something is wrong before conscious thought catches up.
Sometimes early bodily change is informative.
But prediction error is not a truth meter.
A mismatch can arise because the world changed, because the body changed, because the prior was poor, because the sensory signal was noisy, because attention changed, or because the model assigned inappropriate precision.
The existence of mismatch does not establish its cause.
Interpretation remains necessary.
Prediction error and learning
Learning requires some form of updating when expectations fail.
Interoceptive learning can involve repeated exposure to bodily states. Someone who experiences harmless exercise-induced breathlessness may learn that this sensation does not require alarm. Someone with a chronic medical condition may learn that a subtle bodily change reliably precedes a clinically meaningful event.
Those two cases look superficially similar but demand opposite responses.
This is why any practical use of predictive-processing language has to remain domain-specific and evidence-sensitive.
The aim is not to teach people to ignore predictions or sensations.
It is to improve the conditions under which updating becomes possible.
Technology can create new priors
Wearable devices now predict bodily states for us.
A watch says recovery will be poor. A glucose monitor displays a curve before a person notices hunger or fatigue. A health app labels physiological variation as stress.
Those outputs can become priors.
The user may begin to expect tiredness because the score predicts it, or reinterpret an ordinary sensation after an alert.
This does not make wearables harmful. External measurement can improve detection, reveal patterns and support medical care.
It does mean that human interoceptive inference increasingly occurs inside a hybrid system of body, expectation and algorithmic feedback.
Precision is where context enters the model
Prediction error is often described as if the nervous system simply subtracts what happened from what was expected.
Predictive-processing accounts are more complicated because mismatches are also weighted by precision, roughly the inferred reliability or importance of a signal at a particular moment.
That matters because the same physiological change need not have the same inferential impact in every context.
A small increase in heart rate while running for a train is expected. A similar change while lying still at night may receive more attention. Respiratory resistance during deliberate exercise can be interpreted differently from unexpected breathlessness at rest. The sensory event is not read in isolation from context, prior expectation and estimates of reliability.
This does not mean the brain consciously calculates Bayesian probabilities. Precision is a formal property of a family of models. Translating it into ordinary language requires caution.
But the idea helps explain why perception can remain stable despite noisy bodily input and why the same input can sometimes become unusually salient. The system is not merely registering deviations. It is continually determining which deviations should matter enough to alter the current model.
Prediction errors can occur at different levels
“Prediction error” can sound like a single message moving from body to brain.
In computational accounts, errors can be generated and resolved across multiple levels of a hierarchy. Lower-level mismatches may concern sensory features. Higher-level mismatches may concern broader hypotheses about bodily state, emotion, context or action.
This distinction prevents a common conceptual shortcut.
A mismatch at one level does not automatically specify its meaning at another.
A change in cardiac timing does not arrive with the label anxiety. A respiratory sensation does not contain the proposition I am unsafe. Those interpretations depend on higher-order inference, context, memory and learned concepts. Conversely, a high-level expectation can influence which lower-level evidence is attended to or weighted without literally inventing the incoming signal.
This is one reason predictive processing can be clinically suggestive while remaining difficult to prove in any individual case. Several model components can produce similar observable behavior, and different computational descriptions may fit the same data.
Active inference adds action to perception
Some predictive accounts go beyond passive updating.
In active inference, organisms can reduce mismatch not only by changing beliefs but also by acting so that sensory input better matches predicted or preferred states. Moving closer to a heat source when cold is a simple example. Changing breathing, posture, location or social contact can also alter the stream of internal evidence.
That makes interoceptive prediction inseparable from regulation.
The organism is not a spectator waiting for bodily data. It changes the conditions under which data are produced.
This is theoretically important in allostasis, where regulatory systems anticipate needs rather than responding only after variables drift outside acceptable ranges. Yet active-inference language should not be used as a universal explanation for every regulatory action. The framework is powerful partly because it is broad, and breadth creates a corresponding burden to specify the variables, predictions and measurable outcomes in each application.
What would make the evidence stronger
The strongest future evidence will connect computational quantities to independently measurable physiology and behavior rather than treating prediction error as an explanatory label added after the fact.
Useful studies can preregister specific model predictions, manipulate expected bodily states, measure actual physiological change, compare alternative computational models and test whether estimated precision or prediction-error parameters predict behavior out of sample. Replication across cardiac, respiratory and other interoceptive domains would clarify which mechanisms generalize.
Clinical studies need an additional standard. A model that can describe symptoms is not necessarily a model that improves diagnosis or treatment. To support a clinical claim, computational parameters would need sufficient reliability, specificity and incremental value beyond simpler measures.
That distinction matters because predictive language is unusually seductive. It can make a plausible mechanism sound established before the difficult measurement work has been done.
What the evidence does not show
Current evidence does not show that:
- every bodily surprise is an interoceptive prediction error in a measurable computational sense;
- prediction errors are consciously experienced as such;
- one brain region contains or computes all interoceptive prediction error;
- anxiety, depression or another diagnosis is caused by a single prediction-error mechanism;
- high prediction error is inherently bad;
- low prediction error is inherently good;
- bodily mismatch reveals the correct interpretation of an external situation;
- predictive-processing models have replaced other physiological or psychological explanations;
- changing one putative prediction-error signal guarantees clinical improvement;
- active inference is a settled literal description of every level of bodily regulation.
Institute synthesis: update without surrender
Institute synthesis: the useful human capacity is not to eliminate prediction or mismatch.
It is to remain updateable.
A person carries expectations about fatigue, hunger, fear, safety, illness, intimacy and performance. Those expectations shape what becomes noticeable. New evidence arrives from the body and the world. Sometimes the old model deserves to survive. Sometimes it needs revision.
Inner Technology can treat this as a discipline of embodied updating: notice the signal, notice the prior, distinguish confidence from fact, and preserve the possibility that either may need correction.
This is an Institute framing, not a validated computational construct.
From individual prediction to institutional design
Organizations make bodily predictions too.
Work schedules assume how long people can sustain attention. Schools assume when children can sit still. Hospitals assume which symptom reports deserve escalation. Platforms assume which arousal patterns increase engagement.
When those assumptions are wrong, individuals often absorb the mismatch as personal failure.
A richer view asks whether the environment itself is poorly calibrated to human physiology.
Interoceptive prediction error therefore belongs not only to computational neuroscience.
It can also sharpen a social question: when the body repeatedly violates the system's expectations, which model should change?
Editorial limitation note
Predictive-processing language should not be used to tell a person that physical symptoms are “just prediction errors.” New, severe or concerning symptoms may require medical assessment. Computational models describe possible mechanisms of inference; they do not replace physiological diagnosis.
Research references
- Seth AK. Interoceptive inference, emotion, and the embodied self. Trends in Cognitive Sciences. 2013;17(11):565-573. https://doi.org/10.1016/j.tics.2013.09.007
- Barrett LF, Simmons WK. Interoceptive predictions in the brain. Nature Reviews Neuroscience. 2015;16:419-429. https://doi.org/10.1038/nrn3950
- Greenwood BM, Garfinkel SN. Interoceptive Mechanisms and Emotional Processing. Annual Review of Psychology. 2025;76:59-86. https://doi.org/10.1146/annurev-psych-020924-125202
- Toussaint B, Heinzle J, Stephan KE. A computationally informed distinction of interoception and exteroception. Neuroscience & Biobehavioral Reviews. 2024;159:105608. https://doi.org/10.1016/j.neubiorev.2024.105608
- Ventura-Bort C, Weymar M. Transcutaneous auricular vagus nerve stimulation modulates the processing of interoceptive prediction error signals and their role in allostatic regulation. Human Brain Mapping. 2024;45(3):e26613. https://doi.org/10.1002/hbm.26613
- Davey CG. The body intervenes: How active inference explains depression's clinical presentation. Neuroscience & Biobehavioral Reviews. 2025;175:106229. https://doi.org/10.1016/j.neubiorev.2025.106229
- Sterzer P, Keller GB. Predictive processing: Layer-specific prediction error signals in human cortex. Current Biology. 2024;34(10):R496-R498. https://doi.org/10.1016/j.cub.2024.04.031
- Saltafossi M, Heck D, Kluger DS, Varga S. Common threads: Altered interoceptive processes across affective and anxiety disorders. Journal of Affective Disorders. 2025;369:244-254. https://doi.org/10.1016/j.jad.2024.09.135