Predictive Processing and the Brain That Acts on a Model of the Body

Catching a ball requires running on a forecast

Prediction is visible in ordinary skill long before it reaches a clinic. An outfielder starts running before a high ball has arrived anywhere near him, and a practiced one reaches the right place at roughly the right time.

Think about what that requires. Light leaves the ball, enters the eye, and becomes a nerve signal. That signal travels to the back of the head and is worked on by several stages of tissue before anything about it becomes useful. Then a decision travels back out to the legs, and the legs take time to move the body. Add all of that up and you get a delay measured in fractions of a second.

A fraction of a second is a long time for a ball in the air. By the moment the foot lands, the ball is no longer where it was seen. A player who ran to where the ball appeared to be would spend the whole play chasing a ghost of it. The play only works by running to where the ball will be.

Real players run on the forecast

Three psychologists took this seriously enough to test it on real players. Michael McBeath, Dennis Shaffer, and Mary Kaiser wanted to know what rule an outfielder actually follows, so they put cameras on players chasing real fly balls and reconstructed their paths. The players did not compute a landing spot and sprint to it. They ran along curving routes that kept the ball moving in a particular way against the background, a strategy that works because it exploits the geometry of the flight itself.

A sharper image comes from cricket. Michael Land is a visual scientist who studies where the eyes go during real tasks, and with Peter McLeod he filmed the eye movements of batsmen facing balls fired from a bowling machine. He wanted to know where the eyes go while a delivery is in the air. The whole flight lasts about six tenths of a second, and only the fifth of a second after the bounce is much use for deciding the stroke.

The answer was startling. The batsman makes a fast eye movement to the place on the pitch where the ball is about to bounce, and the eyes arrive there before the ball does. In this small group the better batsmen got there sooner, though later gaze studies have not reliably reproduced that particular skill difference. The eyes are not following the ball. They are waiting for it.

The same forecast runs the inside of the body

That is prediction, in a body, doing ordinary work. And the same thing happens on the inside. Your mouth waters before food reaches it. Insulin begins to be released before sugar arrives in the blood.

Physiologists Michael Power and Jay Schulkin gathered this literature and described these anticipatory responses as a class, showing that digestion begins with the sight and smell of a meal rather than with the meal. The body prepares for what it expects to happen, and it prepares in advance because preparing afterwards is too late.

Hold that one sentence, because the whole page hangs on it. A nervous system that only reacts is always behind. A body that must act now has to run on a forecast. Predictive processing is the science of that forecast.

What predictive processing actually claims

Predictive processing claims the nervous system generates its own account of the world from the top down and sends only the mismatch back up. Everything else on this page follows from that reversal.

The older picture of a nervous system is a one-way street. Signals come in from the senses, travel upward through processing stages, and eventually a perception appears at the top. Then a command travels back down to the muscles. Input, then output.

Predictive processing turns most of that traffic around. In this account, the higher levels of the nervous system are continuously generating predictions about what the lower levels should be reporting, and those predictions travel downward. What travels upward is not the raw signal. It is the part of the signal that the prediction failed to account for. That leftover is called prediction error, and it is the currency of the whole system.

Three consequences, each testable

Three consequences follow from predictive processing, and each of them is testable.

First, an expected signal is quiet. If the prediction covered it, there is nothing left to send upward. The system spends its bandwidth on surprise.

Second, perception is a construction. What you experience is the current best model, corrected by error, rather than a direct readout of the world. Your experience is the forecast after the corrections have been applied.

Third, there are two ways to reduce a mismatch. You can change the model to fit the world, which is learning. Or you can change the world to fit the model, which is action. Moving your hand to where you predicted it would be is the second option, which is why this account covers movement as well as perception.

The nervous system does not wait to be told what is happening. It proposes what is happening, and listens for the difference.

The Unified Model of Tone calls this forecasting machinery Prediction. Prediction is the part of the body's organization that is a model of the body rather than the body. It is part of tone rather than tone itself.

Tone is the integrated organization the nervous system maintains across the whole body, and the capacity to move where the moment demands and return to balance afterward. Health is the width of that range. Prediction is one coordinate inside it: how far the system is running on its own account of itself, and how freely that account can be corrected.

Perception as unconscious inference, from Helmholtz to Gregory

Predictive processing began as unconscious inference, Hermann von Helmholtz's 1867 answer to how two flat retinal images become one solid world. The perceiving system settles on the world most likely to have produced its evidence.

Helmholtz was a nineteenth century German physicist who also spent decades measuring the eye. He built instruments to look inside it, measured how fast a nerve conducts, and worked out much of the physics of color and of hearing.

His problem was this. The back of the eye receives a flat, upside down, ambiguous picture. Two of them, in fact, one per eye, slightly different. Out of that you get a stable, solid world of depth and distance that stays put when you move your head. Nothing in the flat image determines that world uniquely. Many different arrangements of objects would produce exactly the same picture.

His answer, developed at length in his 1867 handbook of physiological optics, was that perception must therefore involve a step of reasoning. The system takes the ambiguous image and settles on the arrangement of the world most likely to have produced it, drawing on everything it has learned from experience. He called this unconscious inference, and the word unconscious was the important part. You never get consulted. The conclusion arrives already formed, as a seen world.

It is worth being exact about what he claimed, because it is easy to modernise him too far. Helmholtz framed this as inference drawn from accumulated experience. He did not write it as a formal calculation of probabilities. That framing came much later.

Perceptions are hypotheses

Richard Gregory was a British experimental psychologist who spent his career on visual illusions, and he sharpened the idea in a way that anyone can check at home. His question was why a single unchanging image can be seen in two different ways.

Draw a cube in outline, the figure known as the Necker cube, and stare at it. The front face flips. Nothing on the page changed. Or take a hollow face mask and look at the concave inside from a distance. It refuses to look hollow. It pops out at you as a normal face, and it appears to turn to follow you as you walk past.

Gregory's conclusion, set out in his 1980 paper arguing that perceptions are hypotheses, was that what you see is the system's current best guess about what produced the image. The mask illusion happens because a lifetime of faces has made the outward face overwhelmingly the more likely answer, and that expectation wins against the incoming depth information.

Illusions, in this reading, are not failures. They are the places where the guess is wrong and slow enough that you can catch it happening.

Only the mismatch travels up the nervous system

Prediction error is the only traffic that climbs the nervous system in the predictive processing account. The claim stopped being philosophy in 1999, when a working model of visual cortex made it machinery.

Rajesh Rao and Dana Ballard were computational neuroscientists, which means they built working models of neural circuits and tested whether the models behave the way real tissue does.

They were chasing a specific puzzle in visual cortex. If a neuron is a simple detector for, say, a short line at a particular angle, it should fire whenever that line appears. But real neurons often fire less when the line sits inside a larger pattern that continues smoothly beyond it. Context suppresses them. A detector should not care about context.

So they built a model in which the higher layer sends its prediction down to the lower layer. The lower layer subtracts that prediction from what it is actually receiving, and only the remainder goes back up. Trained on natural images, the model reproduced the puzzling suppression. When the surrounding pattern makes the center predictable, the prediction cancels most of the center's signal, and the neuron goes quiet. The quietness was the point all along.

This is the load bearing teaching idea of the page. What climbs the nervous system is not the world. It is the part of the world the system did not expect.

Rao and Ballard produced a modeling result that accounted for a real effect, and the framework has since been elaborated into detailed proposals about which cells carry predictions and which carry errors. Predictive coding is the leading account of cortical processing. The anatomy that implements it is still being mapped.

Expectation sharpens the signal it quiets

Expectation does not simply turn the volume down on everything. Peter Kok and colleagues, working in Nijmegen, gave people a cue that predicted which image was coming and then measured visual cortex in 2012. The overall response to the expected image was smaller, as predicted. But the pattern of activity became more sharply defined, so the identity of the image could be decoded more accurately. Less total signal, better signal. Expectation edits the incoming stream rather than muting it.

Prediction error is a real, measurable quantity

Prediction error is a measurable physical quantity, recorded from single dopamine neurons in 1997. The cells fire to the gap between forecast and delivery rather than to reward itself.

Wolfram Schultz is a neurophysiologist who spent years with electrodes in the midbrain of monkeys, listening to dopamine neurons. Dopamine had a reputation as the pleasure chemical, and Schultz wanted to know what these cells actually respond to.

The experiment is simple to describe. A monkey learns that a tone is followed by a squirt of juice. Early on, before the animal knows the tone means anything, the dopamine neurons fire when the juice arrives. That looks like a reward signal. Then the animal learns, and something strange happens. The neurons stop firing when the juice arrives. They fire at the tone instead.

The juice has not become less pleasant. What has changed is that it is no longer a surprise. And when the experimenters play the tone and then withhold the juice, the neurons drop below their baseline rate at exactly the moment the juice should have come. A silence where a signal was expected is itself a signal. Schultz, working with Peter Dayan and Read Montague, showed that these cells encode the difference between what was expected and what arrived.

That is a prediction error, in millivolts, in a living animal. Positive when the world beats the forecast. Negative when it falls short. Zero when the forecast was right, no matter how good the outcome was.

Hold that last line. A system running on prediction error is silent when it is correct. Being correct feels like nothing at all.

A movement is a prediction the body makes come true

In active inference, the movement half of predictive processing, a motor command is a prediction the body makes come true. The brain forecasts where the limb will be, and spinal circuitry closes the gap.

Daniel Wolpert is a neuroscientist who studies motor control. In 1995, with Zoubin Ghahramani and Michael Jordan, he tested whether the brain keeps an internal model of the arm. People moved an unseen arm in the dark and then reported where they thought their hand was.

The errors they made had a specific shape over time, and that shape matched what you would expect if the system were running a simulation of the arm and correcting it with sensory feedback. The brain does not simply feel where the arm is. It calculates where the arm should be by now, and uses feeling to keep the calculation honest.

Karl Friston took the next step, and it is the step that matters clinically. Friston is a neuroscientist and brain imaging statistician who spent years looking for a single principle that could cover perception, action, and learning at once. He set it out in a review asking whether one free energy principle could unify brain function.

Any system that persists over time must act to keep surprise low. It has exactly two ways to do that. Change the model to fit the world. Or change the world to fit the model. The second is called active inference.

Rick Adams, Stewart Shipp, and Friston pushed that into the motor system in a paper arguing that the brain issues predictions rather than commands. In their account a motor command is a prediction about where the limb is going to be.

The prediction is wrong at the moment it is made, because the limb is still in the old place. That wrongness is an error signal, and reflex circuitry in the spinal cord resolves it in the only way available, which is by moving the limb until the prediction becomes true.

Read that again slowly, because it inverts the ordinary picture. On this account you do not send an order to your arm. You expect your arm to be somewhere, hold the expectation confidently enough, and the body closes the gap.

What survives the free energy debate

The free energy principle is contested. The philosopher of science Javier Sanchez-Canizares published an assessment separating the good science in the principle from its questionable philosophy in 2021, arguing that its grander interpretations assume much of what they claim to explain. This model does not depend on the grand version.

The surviving claim is narrow and clinical. A mismatch between a body and its own account of itself is expensive. The expense is measurable. And the expense is where this page becomes about patients.

Precision decides which signal the system believes

Precision is the confidence predictive processing assigns to each signal, and it decides whether the prediction or the incoming evidence wins. Most clinically important failures of prediction are failures of precision.

Precision is not accuracy. Accuracy is whether a signal is right. Precision is how much confidence the system assigns to it.

When the prediction and the incoming sense data conflict, the evidence itself does not change. Which one wins depends on the confidence assigned to each, and that is a separate question from what either one reports. The nervous system decides which one to trust, moment by moment.

Turn the confidence up on the incoming signal, and the world wins arguments against your expectations. Turn the confidence up on the prediction, and your expectations win against the world.

Harriet Feldman and Karl Friston made the case that this weighting is what attention actually is. They wrote a 2010 paper recasting attention as the setting of precision on prediction errors. Attending to something means raising the confidence assigned to signals from that channel. Its errors then carry more weight in updating the model. Attention, on this reading, is a gain control on the error stream rather than a spotlight.

This idea reorganises the rest of the page. The system is not usually holding the wrong idea. It is holding an idea with the wrong amount of confidence, or discounting evidence that should have corrected it.

Keep the line clean between precision and gain. Gain is how loudly the system responds relative to the size of the input. Precision is how much the system believes a signal in the first place. They interact constantly, and they are different questions.

Why you cannot tickle yourself

You cannot tickle yourself because the nervous system predicts the touch and cancels it before it is felt. Sensory attenuation is predictive processing demonstrated on your own skin.

Try it. Run your own fingers along your own ribs. Nothing much happens. Have someone else do exactly the same thing and you may not be able to stay still.

Sarah-Jayne Blakemore, Daniel Wolpert, and Chris Frith took this seriously as an experiment. They built a device in which a person moved a lever with one hand and the lever drove a piece of soft foam across the other palm. Then they introduced delays and rotations between the movement and the touch.

Their paper on the central cancellation of self-produced tickle reported the key result: the more the touch departed from what the movement predicted, the more ticklish it felt. A perfectly self-caused touch was least ticklish. Add a fifth of a second of delay and it started to feel like someone else.

The explanation is direct. When the nervous system issues a movement, it also issues a prediction of the sensation that movement will produce, and it cancels the predicted part. What reaches awareness is the leftover. Your own touch is predictable, so most of it is subtracted before you feel it.

Force escalation, measured in newtons

Sukhwinder Shergill and colleagues turned this into an experiment that everyone has effectively already run. Two people sit with their index fingers in a device. One presses on the other's finger. The second person is asked to press back with exactly the same force.

Then the first matches that, and so on. The recorded forces escalated by roughly 38 percent on each turn, because each person felt their own push as weaker than it was and pressed harder to match. Everyone in the exchange is behaving fairly by their own perception, and the force runs away.

That is every argument about who started it, measured in newtons.

The mechanism is also breakable, which is what makes it good evidence rather than a party trick. Shergill's group showed in 2005 that the same cancellation is reduced in people with schizophrenia. Their force matching was closer to correct, because their own touch was attenuated less. A system that fails to cancel its own predicted sensations is a system in which self-produced events start to feel externally caused. That observation has become a central thread in how psychosis is now understood.

Interoception runs on the same predictive architecture

The sense of the body's internal state is built the way vision is. The brain predicts the heart, the lungs, and the gut, and corrects the forecast against what the body reports.

Everything so far has been about the world outside. The clinically decisive half of this page is about the world inside.

Interoception is the sense of the body's own internal condition. It includes the beat of the heart, the fullness of the lungs, the stretch of the gut and the bladder, temperature, the chemistry of the blood, and the state of the tissues. Most of it never reaches awareness. It runs the body regardless.

Lisa Feldman Barrett is a psychologist and neuroscientist, and W. Kyle Simmons is a neuroscientist who works on how the brain represents the body. Their question was whether this internal stream is handled the way vision is. Their answer, set out in a 2015 review proposing that interoceptive signals are predicted rather than merely received, is that the same architecture applies. Deep regions issue predictions about what the body should be reporting, and what travels up is the mismatch.

Anil Seth, a cognitive neuroscientist, carried the idea further in a 2013 paper extending predictive inference to emotion and the feeling of being a body. On his account the felt sense of being alive in a body is itself the running output of interoceptive prediction, rather than a passive reading of visceral state.

This joins directly to the regulation of the body. Peter Sterling, a neuroscientist who spent a career on how nervous systems are built, asked whether the body really defends fixed values or something more forward looking. His 2012 paper setting out allostasis as a model of predictive regulation argues that the body budgets ahead. Blood pressure rises before you stand rather than after you fall. Cortisol climbs in the hours before waking. Insulin moves before the sugar. Regulation is anticipatory, and efficiency is the reason.

That draws one boundary precisely. Set Point is the value the system is defending. Prediction is the forecast the system uses to defend it in advance. Two different questions about the same loop.

The symptom is an inference, and the pain is still real

If a symptom is a perception of the body, and a perception is built from prediction plus evidence, then changing the weighting between them changes the symptom without anything in the tissue changing.

This does not mean the symptom is imagined. That inference is wrong, and it is the single most damaging misreading of this literature. A perception is the real output of a real regulatory system doing the job it was built to do. The pain is real pain. The nausea is real nausea. The dizziness is real dizziness.

What the model claims is something different and more useful, which is that the output depends on the whole inference and not only on the tissue at the far end of it. That is why a person can hurt a great deal with clean imaging, and why another can have alarming imaging and feel nothing.

Surprise is metabolically expensive

Prediction exists because neural signaling is metabolically expensive, and an unresolved prediction error keeps the meter running. The numbers on the brain's energy budget are not intuitive.

The adult brain is roughly two percent of body weight. It consumes something on the order of a fifth of the body's energy at rest. Nothing else in the body is that far out of proportion.

David Attwell and Simon Laughlin, working on the biophysics of neural signaling, wanted to know where that money goes. Their 2001 analysis costing out an energy budget for signaling in grey matter found that the great majority is spent restoring the electrical gradients across cell membranes after signals have passed. Every impulse spends charge, and pumping the charge back is the bill. Signaling is the expense, and the expense is continuous.

Marcus Raichle and Debra Gusnard then asked how much of the budget is triggered by events. Their answer, in a 2002 paper appraising the brain's energy budget, was that the extra energy consumed when a task is performed is small against the ongoing baseline. Most of the spending is intrinsic. The brain is not mostly a responder. It is mostly a system maintaining its own running model, whatever is happening outside.

Unresolved uncertainty is a standing bill

Achim Peters, Bruce McEwen, and Karl Friston made the clinical connection explicit. Peters is a physician who studies brain energy metabolism, and McEwen was the neuroscientist who established how sustained stress reshapes the body. Their 2017 review on why unresolved uncertainty causes disease argues that an unresolved question is itself a sustained energetic demand. A system that cannot settle a mismatch keeps a search process running, keeps stress systems engaged, and keeps paying.

A body carrying a standing prediction error is running an expensive process it never gets to switch off.

That cost has a name. Load is what holding a state costs, and what accumulates when the system cannot stop paying. Prediction is one of the largest reasons a system cannot stop paying.

How a forecast goes wrong

A prediction system fails in four distinct ways: a prior held too strongly, a prior too weak to settle, precision weighted wrongly, and a prediction error the system cannot resolve. Each needs different handling.

One. The prior is too strong

The model is held with so much confidence that incoming evidence cannot correct it. The body keeps producing the predicted experience even when the world has stopped supplying the reason.

Albert Powers and Philip Corlett demonstrated this in a way that is hard to argue with. They took people who hear voices and people who do not, and conditioned everyone by pairing a faint light with a faint tone. Then they showed the light and played nothing.

Their 2017 study showing that conditioned hallucinations follow from overweighted perceptual priors found that people prone to hearing voices reported hearing the tone, confidently, when no tone existed. Modeling their behavior showed the expectation had been given more weight than the ear. The perception was manufactured by the prior.

The clinical version of this is persistent pain, and it gets its own section below.

Two. The prior is too weak

The opposite failure is a model too loose to settle anything. Every signal arrives with full force and demands explanation. Nothing is quiet, because nothing is expected.

This is the state of a system that cannot decide what is normal. Ordinary bodily sensations that should have been cancelled as predictable instead climb to awareness and get interpreted. The heartbeat becomes noticeable. The gut becomes noticeable. Each of them is then a question requiring an answer.

Three. The precision is mis-weighted

This is the most common and the least recognized. The content of the model may be perfectly reasonable. The confidence attached to it is set wrong, in one direction or the other, and often only in one channel.

A body that has learned to distrust its own interior will over-weight threatening interpretations of ordinary signals. A body that has learned to ignore its interior will under-weight signals that should have prompted a correction. Both are precision failures. Neither is a wrong belief in any ordinary sense.

Four. A prediction error the system cannot resolve

This is the failure that matters most to this model, and it is different in kind from the first three.

Normally a mismatch has an exit. Either the model updates, or the body acts and the world changes, and the error closes. But some mismatches have no available exit. The evidence is ambiguous and cannot be settled. The action that would resolve it is unavailable. The signal that would confirm or refute the prediction never arrives cleanly. So the error stays open, and the system keeps carrying it.

What does a body do with an error it cannot close? It holds it. It keeps the relevant channels weighted, keeps the relevant muscles ready, keeps the autonomic setting biased toward the anticipated demand, keeps searching. And because tone is the organization of all of that at once, the unresolved question does not stay in one place. It becomes part of how the whole system is currently arranged.

A standing, unresolvable prediction error held in tone is the model's definition of the distortion.

That sentence is the reason this page exists. In the Unified Model of Tone, the thing every healing discipline is reaching for, under whatever name it uses, is a persistent distortion in how the body registers itself. Stated in the vocabulary of prediction, that distortion is a question the system asked, could not answer, and never put down.

It shows up as tension held without a task, as regulation biased without a demand, as a perception maintained without its cause. The person does not hold it as a belief. The organization holds it as a shape.

Pain tracks the inference, not the tissue

Persistent pain is the clearest clinical case of prediction. Disc degeneration appears in 37 percent of pain free 20 year olds and 96 percent of pain free 80 year olds, so pain tracks the nervous system's inference rather than the tissue.

Begin with the imaging, because it removes an assumption most people arrive with.

Waleed Brinjikji, a neuroradiologist, led a team asking a simple question. What do the spines of people with no back pain at all look like on a scan? In 2015 they pooled 33 studies covering more than 3,000 pain free adults.

Their review of imaging features of spinal degeneration in people without symptoms found disc degeneration in 37 percent of pain free 20 year olds. By age 80 it was 96 percent. Disc bulges were present in 30 percent at age 20. These findings behave like grey hair. They accumulate with age, and their presence says little about whether a given person hurts.

Pain and tissue state are loosely coupled. That is the finding, and it is not controversial. What has been missing is an account of what pain is tracking instead.

The imprecision hypothesis

Lorimer Moseley is a pain scientist, and Johan Vlaeyen is a psychologist who studies fear and avoidance in pain. In 2015 they proposed what they called the imprecision hypothesis of chronic pain.

Their argument is that the nervous system learns which patterns of input signal danger to body tissue, and that in chronic pain this learning becomes imprecise. The category widens. Inputs that never meant danger start to fall inside it. The system is not malfunctioning. It has learned a bad generalisation and is applying it faithfully.

Abby Tabor and colleagues put the same argument in formal terms in a paper treating pain as a statistical account of bodily danger. In their framing pain is the system's best inference about threat to the body, assembled from prior expectation and current evidence. When the prior is strong and the evidence is ambiguous, the inference can be maintained long after the tissue has healed, because nothing arriving is decisive enough to overturn it.

Katja Wiech, a neuroscientist working on pain and cognition, surveyed how far this reaches in a review deconstructing the sensation of pain. Expectation, attention, context, and what a person believes the sensation means all change the pain that is felt, and they do so by changing the inference rather than by changing the tissue.

A loop that feeds on its own output

Marieke Jepma and Tor Wager then found the mechanism that makes this self-sustaining, and it is the finding that should be better known. They gave people expectations about how much heat would hurt, delivered the heat, and tracked both the reported pain and how the expectation changed afterwards.

Their 2018 work demonstrating self-reinforcing expectancy effects on pain showed a closed loop. The expectation shifted the experienced pain toward itself. The person then learned from the experience they had just had, which confirmed the expectation. Round it goes.

Read that carefully. A loop like that needs no ongoing input from tissue to keep running. It is fed by its own output.

This is not confined to pain. Peter Henningsen and colleagues extended the same reading to persistent physical symptoms generally, in a paper describing them as perceptual dysregulation. Symptoms that resist explanation are better modeled as a failure of the inference than as a hidden lesion nobody has found yet. Giulio Ongaro and Ted Kaptchuk made the same point from the placebo side in an editorial on symptom perception and the Bayesian brain.

None of this changes the workup. Some pain does come from tissue that needs attention, and some unexplained symptoms turn out to have a cause that testing can find. A prediction reading of a symptom is an addition to a proper workup and never a replacement for one.

Placebo and nocebo are the prediction system measured

In 2011, expectation roughly doubled the analgesia of a running opioid infusion or abolished it, with the drug constant throughout. The forecast reached the same physiology the opioid did.

The placebo effect is often described as if it were a courtesy, a way of saying a person felt better because they wanted to. The physiology says otherwise, and the experiments are unusually clean.

Jon Levine, Newton Gordon, and Howard Fields were working with patients after dental surgery in the 1970s, and they wanted to test whether placebo pain relief involves the body's own opioid system. They gave naloxone, a drug that blocks opioid receptors. Their 1978 trial on the mechanism of placebo analgesia found that the relief produced by placebo was abolished by the blocker. Whatever was happening was chemical, and it was the body's own chemistry.

Fritz Eippert and colleagues repeated that logic inside a scanner three decades later. Their 2009 study showing that the opioid descending pain control system underlies placebo analgesia found that blocking opioid receptors removed both the reported relief and the brain activity that normally accompanies it. Tor Wager's earlier imaging work on brain changes during the anticipation and experience of pain under placebo had already shown that the shift begins during anticipation, before the painful stimulus arrives. The forecast is doing the work.

The forecast was worth as much as the drug

Ulrike Bingel ran the experiment that makes the point unmissable to a clinician. She is a neurologist who studies pain and expectation. She gave patients remifentanil, a genuinely powerful opioid, at a constant infusion, and manipulated nothing except what they were told. Her 2011 study on the effect of treatment expectation on drug efficacy found that positive expectation roughly doubled the analgesia. Negative expectation abolished it entirely while the drug was still infusing at the same rate.

The drug did not change. The prediction did. And the prediction was worth as much as the drug.

Fabrizio Benedetti's group generalised this with a decisive design. Luana Colloca and Benedetti compared giving a treatment openly, with the patient watching, against giving exactly the same dose hidden by an automatic pump. Their 2004 review of overt versus covert treatment for pain, anxiety, and Parkinson's disease reported that hidden doses were reliably less effective across all three. A known dose and an unknown dose are not the same intervention, because the known one includes a prediction.

The effect is not limited to pain. Raul de la Fuente-Fernandez and colleagues studied people with Parkinson's disease, a condition in which dopamine producing cells are lost. They gave a placebo and scanned for dopamine release. Their 2001 study measuring dopamine release after placebo in Parkinson's disease found substantial release of the patients' own dopamine in the striatum, comparable to what an active drug produces. An expectation of benefit released the chemical the disease is short of.

Luana Colloca and Arthur Barsky wrote the clinical summary of this field in a review of placebo and nocebo effects, and their framing is the useful one. These are not artefacts to be subtracted out of trials. They are the measurable output of the prediction system, and it is running in every encounter whether anyone attends to it or not.

Nocebo is the same machinery pointed the other way

Nocebo is the prediction of harm, and it produces real symptoms by the same route.

The cleanest demonstration is a statin trial. Frances Wood and colleagues at Imperial College ran a study called SAMSON, in which patients who had stopped statins because of side effects took twelve one month bottles in random order. Four contained a statin, four contained placebo, and four were empty. Patients rated their symptoms daily.

The 2020 result of the trial comparing statin, placebo, and no tablet was that symptom scores on placebo months were almost as high as on statin months, while no tablet months were much lower. The great majority of the symptom burden was attributable to taking a tablet rather than to the drug in it.

An independent team reached compatible conclusions at larger scale. In 2021 Emily Herrett and colleagues ran two hundred individual randomised trials of statin against placebo in primary care and found no overall difference in muscle symptoms between the two.

The symptoms were not invented. They were reported daily and scored, and they were real experiences. Nothing on this page is advice about any medication, and it is no reason to stop one. What the trials show is that a prediction of harm, held about a tablet, produced most of the harm attributed to the tablet. That is the nocebo mechanism doing exactly what predictive processing says it does.

How Prediction is measured, and what each instrument misses

Prediction is read through four instruments: expectancy paradigms, mismatch negativity, sensory attenuation, and interoceptive accuracy. Each reads one channel, and each carries a boundary that is part of the reading.

Expectancy paradigms

Tell one group a treatment will help and another that it will not, hold everything else constant, and measure the difference. Bingel's remifentanil study is the sharpest example, because the drug is constant and only the sentence changes. The strength of the design is that it isolates prediction from pharmacology. The limit is that it measures the effect of a manipulated expectation in a laboratory, which is not the same as measuring the expectation a person walks in carrying.

Mismatch negativity

Play a person a repeating tone, then slip in a different one, and record the electrical activity at the scalp. About a fifth of a second after the odd tone, a distinct negative deflection appears. It shows up even when the person is reading a book and paying no attention, and even during sleep.

Marta Garrido and colleagues reviewed the mechanisms behind it in a 2009 paper assessing what generates the mismatch negativity, and the leading account is that it is a prediction error signal from auditory cortex. Its value is that it reads the system's model of regularity without asking the person anything at all. Its limit is that it reads one channel, hearing, at one timescale, and generalising from it to the whole of a person's predictive style overreaches.

Sensory attenuation

The force matching task is a working clinical measure. How much weaker does a self-generated force feel than an externally applied one of the same size? That number quantifies how strongly the system cancels its own predicted sensations, and it is reduced in schizophrenia. Its limit is that it reads the motor to sensory prediction loop specifically, and other predictive functions can be intact when it is not.

Interoceptive accuracy

Ask a person to count their own heartbeats without taking a pulse, and compare against a recording. Sarah Garfinkel and colleagues showed in a 2015 paper separating interoceptive accuracy from interoceptive awareness that these come apart. Some people are accurate and do not know it.

Others are confident and wrong. The gap between the two is itself informative, and it reads precision directly. The limit is that heartbeat counting is a crude and contested task, and it does not straightforwardly generalise to the gut, the lungs, or the tissues.

No single instrument reads the whole dimension

There is no single instrument that reads Prediction the way a cuff reads blood pressure. What exists is a set of partial windows, each valid in its own channel. That is the ordinary condition of a young measurement science, and it is the reason prediction is read best by combining what several instruments say.

Conditions where Prediction carries most of the weight

Prediction carries most of the weight in persistent pain, fibromyalgia, anxiety, unexplained symptoms, tinnitus, and much of mental health. Every condition in this library is a combination of dimensions, and each pairing here has a specific reason.

It also carries real weight in long covid. It carries weight in trauma, where a prediction of danger outlives the danger, and in gut health, where interoceptive prediction meets visceral sensation. And in sleep, the anticipation of not sleeping is itself part of what prevents it.

How Prediction relates to the other dimensions of tone

Prediction is one of the foundational dimensions of tone, and it stays useful only while the lines between it and its neighbors stay sharp. Here is where prediction ends and each neighbor begins.

Set Point is the value the system is defending. Prediction is the forecast it uses to move that value ahead of demand. Allostasis is where they meet, and they remain two questions. What is being defended, and on what forecast.

Gain is how loudly the system responds relative to the size of the input. Prediction is what the system expected the input to be. A raised gain amplifies everything arriving. A strong prior changes what counts as arriving at all.

Oscillation is the rhythm and range a single system moves through. Prediction runs on those rhythms and is not one of them.

Coupling is whether separate systems stay in step with one another. Oscillation is within a system, coupling is between systems, and prediction is a property of the model either of them is running.

Load is what holding a state costs and what accumulates. An unresolved prediction error is one of the most reliable ways a body ends up paying continuously. Prediction generates the demand. Load is the bill.

Constraint and Slack is where the body has room to move, mechanically and neurally. It matters here because active inference requires movement. A prediction that cannot be tested by moving is a prediction that cannot be corrected.

Input Quality is the fidelity of the signal coming in. This is the closest neighbor and the line must be exact. Input quality is how good the evidence is. Prediction is what the system does with evidence of whatever quality. Degrade the evidence and the prior necessarily carries more of the result, which is why the two are so often found together.

Time Course is how old the problem is. Prediction is the part of tone most changed by age, because every cycle of a self-confirming loop makes the prior stronger. A recent expectation is a suggestion. A decade-old one is architecture.

The inputs and interventions that correct a prediction

A prediction corrects when the system receives evidence it cannot dismiss. That is the whole logic of correcting a prediction, and it explains why the interventions that do it are so varied.

Evidence the prior cannot explain away

Yoni Ashar and colleagues tested a treatment built directly on this logic. Pain reprocessing therapy works by helping a person attend to a familiar pain while receiving evidence, from movement and from the therapist, that the sensation is not signaling danger.

They ran a randomised trial, published in 2022, comparing pain reprocessing therapy against placebo and usual care in chronic back pain. At the end of treatment, sixty six percent of the treated group were pain free or nearly pain free. The figures were twenty percent on placebo and ten percent on usual care.

The boundaries are specific. This is a single trial with a selected population, and the participants had relatively low disability and no serious structural findings. It is a strong result in a well defined group rather than a general claim about all back pain.

Placebo works even when labeled as placebo

The strangest finding in this area is that placebo works when you say it is placebo. Ted Kaptchuk and colleagues gave patients with irritable bowel syndrome pills openly labeled as placebo, explained honestly what they were, and explained that the body can respond to the ritual.

Their 2010 trial of placebos given without deception found significantly better symptom relief than no treatment. Claudia Carvalho's group replicated the design in chronic low back pain, in a 2016 randomised trial adding open-label placebo to usual care, and reported reduced pain and disability.

These are modest effects in unblinded trials and they should be held at that size. But they matter conceptually, because a prediction does not require a false belief. It requires a plausible expectation attached to an event the body registers.

Movement

Movement is the most underrated way to correct a prediction, and the reason is structural. Active inference says a prediction is tested by acting. A body that moves generates a continuous stream of evidence about what it can do, delivered in the only format that can correct a motor prediction. A body that stops moving to protect itself removes the evidence that would update the forecast, and the forecast then hardens without opposition.

Attention to the interior, trained

Slow breathing, interoceptive attention practices, and any structured contact with the body's own signals work on precision rather than on content. They change how much the system trusts what the body is reporting. That is the lever, and it is a different lever from persuasion.

Where the evidence is mixed

Explaining the biology of pain to a patient is a prediction intervention, and the evidence for it is genuinely mixed. Adriaan Louw and colleagues reviewed the trials in a 2016 systematic review of pain neuroscience education. They reported benefit for pain, for pain beliefs, for function and for the fear of movement, and the effects were strongest when the education was paired with movement rather than delivered on its own. Later trials have been less consistent on pain intensity itself.

The model expects exactly that pattern, and this is where it earns its keep. An explanation is information offered to a system. Whether it lands depends on the tone that receives it. Where a person's forecast is already loosely held, a good explanation can move it in one conversation.

Where the prior has been confirmed by a thousand painful movements, a sentence is a very small piece of evidence against a very large body of it. Input meets tone. The same input, the same words, different results, and the variation is the signal rather than the noise.

The distortion is a standing prediction error held in tone

The Unified Model of Tone reads prediction as the dimension in which the body carries an account of itself, and a standing unresolvable prediction error as the distortion every healing discipline is reaching for.

Everything above is established science, argued by the people who did the work. What follows is this model's own reading of it, stated as such.

The Unified Model of Tone holds that a body is one organization, expressed at every scale at once, and that health is the width of the range that organization can move through and return from. Prediction is the part of that organization that carries the body's account of itself.

The model's specific claim is this. A question the system asked, could not answer, and never put down becomes part of how the whole body is arranged. The distortion is neither a structure out of place nor a belief awaiting correction. It is a standing prediction error held in tone.

That reframes what a good intervention is doing. It delivers evidence in a form the system cannot dismiss, at the place where the question is being held, so the body can close the loop and stop paying. Persuasion has very little to do with it.

Where that place is matters more than how hard anything is done. The place a symptom shows up is usually not where the deepest driver sits. A system carries several points of critical tension at once, in varying degrees of potential, and patterns layer. A presenting pattern often sits on top of a deeper driver, and when the driver releases the surface pattern lets go with it.

The art in every discipline is finding the point that yields the most change for the least force at that moment. Specificity means correspondence, matching the right input to the actual state. It does not mean force. Magnitude is a separate axis entirely, running from the lightest contact through to surgery.

The bidirectional test: restoring versus pushing

One trial design separates an intervention that restores regulation from one that pushes an output.

The model predicts bidirectional restoration. A genuine tonal correction should move a dysregulated loop toward the middle from either side. A person running too high should trend down. A person running too low should trend up. Same input, opposite directions, converging on the middle.

A drug does not do this. An opioid pushes analgesia one way in everyone. A stimulant pushes arousal one way in everyone. Direction is a property of the agent.

Here the model's claim is specific and testable. Where prediction is dominant, the same intervention should reduce an over-weighted prior in one person and firm up an under-weighted one in another, and both should move toward the middle.

Measure sensory attenuation, mismatch responses, and interoceptive accuracy in a group before and after, split by baseline. If everyone moves the same direction regardless of where they started, the effect is a push. It helps whichever group it happens to point at and carries the other group further from the middle.

Restoring versus masking

Quieting a signal and resolving what generates the signal are different aims, and both are legitimate. Symptomatic relief has real value. Acute medicine saves lives, and nothing here argues otherwise. Nobody should change a treatment because of a page on the internet.

The distinction is about mechanism and aim. An intervention that lowers a symptom without touching the standing error leaves the error running, and running errors cost energy. An intervention that helps the system close the question removes the demand itself.

The reason this distinction has been so hard to see is that tone has never been recognized as its own regulatory system. Practitioners across every field are already working with it. They simply have not had a name for it, or a way to say which of the two things they were doing.

What the composition claim adds

The components are not claimed as new. Helmholtz, Gregory, Rao and Ballard, Schultz, Friston, Barrett, Moseley, and Kaptchuk did the work, from 1867 to the 2020s, and they are credited above. The claim is about composition.

Prediction, set point, gain, oscillation, coupling, load, constraint, input quality, and time course are one organization read through different instruments, and a condition is a specific combination of them. A relabelling makes no new predictions. This one predicts bidirectional restoration, which no collection of independent mechanisms expects, and it names the measurements that read it.

That is the difference between a name and a model.

Frequently asked