Predictive Processing: The Brain Guesses First and Checks the World Later

Introduction

There is a mask in a lot of psychology labs. It is an ordinary theatrical face mask, moulded plastic, nothing special about it. It hangs on a slow turntable under a light, and as it rotates you watch the front of the face swing past, then the edge, then the hollow inside.

Except you do not see the hollow inside.

When the concave back of the mask comes round, it flips. It pops out at you as a normal face, nose forward, cheeks rounded, sitting there solidly in the wrong direction. You know it is hollow. You watched it turn. It makes no difference at all. The illusion is so stubborn that people reach out and touch the mask to check, and it still looks convex while their finger is inside it.

This is called the hollow mask illusion, or more formally a depth inversion illusion, and it belongs to a large family of tricks in which what you know loses to what you expect. If you want the wider tour of that family, we have written about how optical illusions work separately. What matters here is one detail.

Some people are not fooled by it.

In a 2025 study, seventy seven people with schizophrenia, fifty healthy control participants and forty people who hear voices but have no psychiatric diagnosis all sat in front of a hollow mask task [1]. The control group reported the illusion significantly more often than the patients did. Earlier work found the same pattern in young people at high risk of psychosis, with forty four at-risk participants and twenty nine controls, and the at-risk group again saw the mask more accurately [2].

Read that again, because the direction is the surprising part. The patients were not more deceived. They were less deceived. They looked at a rotating hollow mask and reported a rotating hollow mask, which is what is physically there, and which almost nobody else manages to do.

Being harder to fool is the symptom.

That result only makes sense if you accept something strange about perception. It means seeing is not a recording. It means the convex face you see is not arriving from the mask. It is being supplied by you, from the inside, and the mask is only voting on it. When that internal supply is weakened, the vote from the mask starts winning, and you see what is actually there.

The framework built around that idea is called predictive processing. Its central claim is that the brain runs ahead of its senses. It is constantly generating a best guess about what the world is doing, sending that guess down through the perceptual system, and using the incoming signal not as the picture but as a correction to the picture.

This article is about that claim. Where it came from, what it explains beautifully, and where the last five years of evidence have made a lot of careful people uneasy about it.

Because here is the thing you will not find in most explanations of predictive processing. It is not settled. Popular accounts tend to present it as the modern consensus about how the brain works, and then stop. Meanwhile, in the actual literature, researchers have been publishing papers with titles like "Is predictive coding falsifiable?" and "Limited Evidence for Sensory Prediction Error Responses in Visual Cortex of Macaques and Humans" [3] [4]. That argument is the most interesting thing about the field right now, and leaving it out would be a kind of lie.

So we will do both. The idea, taken seriously. Then the case against it, taken just as seriously.

The Guess Comes First

Start with what the senses actually deliver, because it is much worse than it feels.

Your retina is a curved sheet of cells that receives a flat, upside down, low resolution projection of a three dimensional world. It has a hole in it where the optic nerve leaves. Your eyes jump three or four times a second, and during each jump the image smears and the input is largely suppressed. Colour vision is concentrated in a patch about the size of your thumbnail at arm's length. Everything else is coarse.

None of that is your experience. Your experience is a stable, wide, coloured, three dimensional room with no hole in it.

Something is doing an enormous amount of work in between, and it is doing it fast. Visual scenes get categorised in around a seventh of a second, which we have covered in more detail in our piece on how fast the brain recognises a visual pattern. That speed is itself an argument. A system that waited for signals to climb all the way up a processing hierarchy before deciding anything would be slower than that.

The person who first said the obvious thing about this was Hermann von Helmholtz, in the 1860s, without a single measurement of a neuron. In his Handbuch der physiologischen Optik, published by Leopold Voss in Leipzig from 1867, he argued that perception is a form of unconscious inference. The senses give you effects. You need causes. Getting from effect to cause is a guess, and your visual system makes that guess so quickly and so automatically that it does not feel like a guess at all. It feels like looking.

Helmholtz had no way to test this. He had the argument, and the argument was correct, and then the field spent about a hundred and thirty years catching up to him.

Where the Guesses Come From

If perception is a guess, the obvious next question is where the guess comes from. It is not arbitrary and it is not conscious. It is built from statistics you never chose to collect.

Take one you are using right now. Light comes from above. It has done for the entire history of life on this planet, and your visual system has absorbed that fact so thoroughly that it uses shading to infer shape without ever consulting you. Print a page of identical circles, shade half of them dark at the top and half dark at the bottom, and one set will look like bumps and the other like dents. Turn the page upside down and they swap. Nothing on the paper changed. The assumption did.

Or take the one running the hollow mask. Faces stick out. Every face you have seen since infancy has stuck out, which is an unusually consistent training set, and the resulting expectation is unusually hard to shift.

That is what a prior is. Not a belief you hold, but a bet the system has already placed on your behalf, weighted by how often it has paid off. Some are old enough to be built in. Others are learned within minutes, which is why a room stops looking strange after you have been in it a while and why a new accent becomes easy to follow after a few hours of listening.

This also explains something that annoys people about illusions. Knowing about an illusion almost never dissolves it. You can read a full account of why the mask flips, understand every word, and it will still flip. That is because the knowledge lives at a level of the system that does not get a vote in the fast local inference happening several floors below it. Your understanding and your perception are running on different timescales, using different evidence, and the fast one gets to decide what you see.

Which is worth remembering the next time you are certain about something you looked at.

Why You Never Catch It Happening

Here is the part that makes this hard to believe from the inside. If your brain is doing all of this, why does looking feel like looking?

The answer is that you have no access to the process, only to the result. The comparison between prediction and input happens below the level anything reports to you. What reaches awareness is the settled interpretation, already cleaned up, already committed. You do not experience a draft being corrected. You experience the finished page.

You can catch the machinery only at the edges, when it fails or when it takes unusually long. A word misheard in a noisy room and then suddenly heard correctly, without the sound repeating. A shape in low light that resolves into a coat on a chair, after being something worse for half a second. The moment a piece of music you have never heard before stops sounding random. Each of those is a model updating, and the reason you notice is that the update was slow enough to have a before and an after.

The rest of the time the correction is fast enough that there is no gap to notice, and perception presents itself as direct contact with the world.

This is also why the framework is genuinely difficult to argue about. Nothing in your experience distinguishes seeing from an extremely well corrected guess. The two feel identical from the inside, which means the question has to be settled with electrodes and scanners rather than introspection. Everyone doing this research is working against the strong intuition that they are simply looking at things.

Three Ideas That Keep Getting Called One Thing

Before going further, there is a piece of vocabulary that causes an enormous amount of confusion, including in places that ought to know better. Three different ideas circulate under names that sound interchangeable. They are not interchangeable, and mixing them up is the single most common way to get this topic wrong.

Predictive coding is an algorithm. It is a specific proposal about how signals move through a hierarchy: higher levels send predictions down, lower levels send back only the part of the signal that was not predicted. That is a narrow, testable engineering claim about wiring. It was given its modern computational form by Rajesh Rao and Dana Ballard in 1999 [5], building on an architectural proposal published by Mumford in 1992 [6]. Confusingly, the same phrase also means something completely unrelated in signal processing, where predictive coding is a family of compression techniques for video and audio. Search for the term and you will get both.

Predictive processing is a framework. It takes the prediction error idea and applies it well beyond sensory wiring, to attention, action, emotion, memory and mental illness. Andy Clark's 2013 target article "Whatever next?" is the piece that made this framing mainstream [7].

Two down. The third is the one that causes the arguments.

The free energy principle is broader still, and this is where people get lost. Karl Friston's claim is not really about neurons at all. It is a claim about self organising systems in general: any system that maintains itself against a changing environment must act to minimise a quantity called variational free energy, which under simplifying assumptions behaves like prediction error [8] [9].

Predictive coding falls out of it as a special case under a particular approximation [10]. That generality is either the theory's great strength or its central problem, depending on who you ask, and we will come back to that fight. A clean modern statement of how the three fit together is available if you want the formal version [11].

It is worth being slightly annoyed about this. A field that cannot keep its own three central terms apart has made its ideas harder to evaluate than they need to be, and the resulting confusion is not only the public's fault.

Keep the three separate as you read. An experiment can undermine predictive coding as an algorithm while leaving the broader framework standing. And the free energy principle can survive almost anything, which is exactly what worries its critics.

How the Loop Is Supposed to Work

The mechanism is easier than the vocabulary.

Picture the perceptual system as a stack of levels. Low levels deal in edges, contrast, brief sounds. Higher levels deal in objects, faces, words, situations. In the old textbook picture, information flows upward: edges become shapes, shapes become objects, objects become meaning.

Predictive coding turns most of that around. Each level holds a model of what the level below it should be reporting, and it sends that expectation downward. The lower level compares what it actually got against what it was told to expect. If they match, almost nothing travels upward. If they do not match, the difference travels upward, and that difference is the prediction error.

So the upward channel is not carrying the world. It is carrying the news. Everything the system already anticipated is subtracted out before it climbs, which is why the arrangement is efficient in the first place: you never pay to transmit what was already known.

There is an anatomical reason people find this attractive. The connections running backwards and sideways through cortex vastly outnumber the ones running forwards from the senses. Something is being sent down there, in bulk, all the time. Our overview of how the visual cortex is organised covers that layered structure in more detail. A 2018 review argued that this predict-and-subtract arrangement is a canonical cortical computation, meaning the cortex does roughly the same trick everywhere rather than a different trick per region [12].

The strongest human evidence for the hierarchy being genuinely layered came from an EEG and MEG study that built expectations at two levels at once, so that a surprise at one level could itself be expected at the level above [13]. Work in primates has since traced hierarchical prediction and prediction error across large scale cortical networks [14].

Notice the two exits from prediction error at the bottom of that loop. You can change the model to fit the world, which is perception and learning. Or you can change the world to fit the model, which is action. Reaching for a cup is, in this account, the same operation as recognising a cup, run in the other direction. That symmetry is the part of the framework people find hardest to swallow and also the part that makes it ambitious.

1999: The Model That Made It Testable

Helmholtz gave the field an argument. What it needed was a mechanism that made a wrong prediction of its own, and that arrived in 1999.

Rao and Ballard built a hierarchical network that learned to predict natural images, with feedback carrying predictions and feedforward carrying residual error. Then they looked at what their model neurons did, and found they reproduced a set of effects that had been puzzling visual neuroscientists for years: extra-classical receptive field effects, where a neuron's response to a stimulus is changed by things happening outside the region it is supposed to care about [5].

That was the moment the idea stopped being philosophy. A neuron that responds less when a bar is extended beyond its receptive field looks arbitrary if you think of neurons as feature detectors. It looks obvious if you think of them as error reporters, because a longer bar is more predictable from context, so there is less residual to report.

The idea has kept moving. Rao's own recent work extends it to sequence learning and to models that act as well as perceive [15] [16].

Other groups have built hybrid schemes that combine fast and slow inference [17]. There is an active argument about whether the relevant dynamics behind all this are oscillatory or transient [18]. That sounds like a technicality and is not one. The two accounts disagree about when in a response you should expect to find the error signal, which means they disagree about what a failed search for it would prove.

That last stretch of the timeline is the part popular explainers tend to skip. Hold onto it.

The Sound That Never Came

Of all the evidence for this framework, one result is the hardest to explain any other way.

Play someone a steady rhythm of tones. Then leave one out. Not a wrong tone. Nothing. Silence where a sound should have been.

The brain responds. There is a measurable neural response to the absence, timed to when the missing sound would have arrived [19] [20] [21].

Sit with how odd that is. A response requires something to respond to, and nothing happened. The only way to generate a signal at the moment of a non-event is to have already committed to an expectation about that moment. The prediction has to exist in advance, or there is nothing for the silence to violate.

There is a version of this you have probably experienced. A fridge or an air conditioner runs in the background for hours and you stop hearing it entirely. Then it switches off, and the silence is loud. Nothing arrived at your ears at that moment. What you noticed was your own prediction failing to be met, which is the only thing available to notice.

The related and much older finding is the mismatch negativity, an electrical response that appears when a sound breaks an established pattern, and which shows up whether or not you are paying attention. It has a natural reading as a prediction error signal, and a detailed neuronal model has been built on that reading [22].

The complication, and it is a serious one, is that neurons also just get tired. A repeated stimulus produces a shrinking response through simple adaptation, with no prediction involved. So a smaller response to the expected and a bigger response to the unexpected could be prediction, or it could be fatigue. Distinguishing these is a whole research programme in itself. One influential study separated them in time, showing that repetition suppression and expectation suppression are dissociable in early auditory responses [23], and further work has examined how context shapes repetition effects [24].

Remember that confound. It comes back later and does real damage.

Attention Is a Volume Knob

If everything in the model is a guess, the system needs some way of deciding how much to trust each guess and each error signal. In this framework that quantity is called precision, and it is roughly the expected reliability of a signal.

A prediction error from a source you consider reliable gets amplified and shapes the model strongly. One from a source you consider noisy gets turned down and changes almost nothing. Precision is the gain control.

The claim that follows is that attention is precision weighting. Attending to something is not shining a light on it. It is deciding in advance that errors from that source are worth listening to, and turning up their volume.

That reframing makes a specific prediction, and it is a strange one. Expectation should quieten the response to a stimulus, since expected things generate little error. Attention should do the opposite to the same stimulus. So attending to something you expect should reverse the quietening. That is exactly what one study found: attention reversed the silencing effect of prediction on sensory signals [25]. Two effects that both change how strongly a neuron responds, pulling in opposite directions, from the same mechanism.

Related work has looked at where the gain control might live anatomically, with the pulvinar as a candidate [26], and at the conceptual relationship between attention and conscious perception in a prediction based system [27].

Hearing has been a productive place to test all of this, because auditory expectations can be set up and broken on a precise clock in a way that vision does not easily allow [28] [29]. That same gain control is the hinge between attending to something and actually retaining it.

This is the point where the framework starts to feel less like a theory of vision and more like a theory of everything, which is a feeling worth paying attention to. Keep it in mind for later.

Less Signal, Better Picture

Here is a result worth slowing down for, because it looks like a contradiction until it does not.

When you expect a particular visual stimulus, primary visual cortex responds less to it. Lower amplitude, weaker signal. On a simple view that should mean worse perception, since less activity ought to mean less information.

The opposite happened. Using multivariate pattern analysis, researchers found that while expectation lowered the amplitude of the V1 response, it improved the representation of the stimulus in that same region, and the improvement tracked how much better people performed [30].

Less activity. Better information. The paper is titled "Less Is More" and it earns the title.

This is what the framework predicts if expectation is doing its job. If the model already accounts for most of the input, the leftover activity is the informative part, and cleaning away the predictable bulk sharpens what remains rather than degrading it. Later work traced how long term priors reach back into perception through long range feedback connections [31].

Hold that finding loosely for now. There is a rival explanation, and it arrives in a few sections.

The Mask That Fools Almost Everyone

Now back to the mask, with the machinery in place.

Why does a hollow face pop out at you? Because you have spent your entire life looking at faces, and every one of them stuck out. The prior that faces are convex is about as strong as a visual prior gets. When the mask is concave, the stereo and shading evidence says hollow, the prior says convex, and the prior wins so decisively that you cannot override it by knowing better.

If perception were a recording, this could not happen. A recording of a hollow object is a recording of a hollow object. The illusion exists because the thing you see is the model's best guess, and the guess has been overwhelmingly trained in one direction.

Which sets up the test. If the illusion depends on priors beating evidence, then anyone whose priors carry less weight relative to the evidence should be less fooled. And the predictive coding account of psychosis proposes exactly that: that in schizophrenia the balance between top down expectation and bottom up sensory evidence is disturbed [32] [33]. An older strand of the same argument describes it as a connection problem rather than a regional one, with the failure showing up in how areas talk to each other and in the brain's ability to tell its own signals from the world's [34].

The prediction is counter-intuitive and it holds up. Across a range of studies, people with schizophrenia are less susceptible to depth inversion illusions than controls [35] [36] [37].

That is a number of separate laboratories arriving at the same backwards result, which is the kind of thing that makes a finding worth trusting. A systematic review pulled together forty five studies of visual illusions in schizophrenia and found concordant evidence of abnormal illusion processing, with facial depth inversion among the clearest cases [38].

It is rare for a theory of perception to make a prediction this specific and this easy to get wrong. If the account were backwards, patients would be more fooled rather than less.

There is also a mechanism study. Using dynamic causal modelling on functional imaging data, researchers examined what changes in the brain when patients view hollow faces, and found strengthened bottom up and weakened top down effective connectivity in the patient group, while controls strengthened their top down influence when looking at the same stimulus [39]. In a group at ultra high risk for psychosis, more accurate perception of the mask went with weaker connectivity within the fronto-parietal network [2]. Another study connected illusion performance to cortical thickness in frontal and parietal regions in twenty two first episode patients who had never taken antipsychotic medication [40].

None of that tells you why the effect exists, only where it shows up. Effective connectivity is a model fitted to data rather than a wire anyone can point at, and it is worth keeping the distance between those two things in view.

That first row is not a typo, and it is not there for balance. It is there because it is true.

What the Mask Does Not Prove

A study of forty two people at clinical high risk for psychosis and forty four healthy controls tested binocular depth perception directly and found no dysfunction at all. Performance matched the control group, and it did not correlate with positive symptoms [41]. The authors concluded that impaired binocular depth perception should not be considered a marker in that group.

The 2025 study with the largest sample found something similarly deflating for anyone hoping this becomes a clinical tool. It compared people with schizophrenia, healthy controls and non-clinical voice hearers, and although controls saw the illusion more often, there was no consistent association between illusion perception and symptom severity in either group that experienced unusual perceptions [1]. Whatever the mask is measuring, it does not track how ill someone is.

None of this is a reason to discard the finding. It is a reason to state it at the size it actually is.

The systematic review is blunt about the state of the literature too, noting significant methodological disparities across the forty five studies and concluding that the usefulness of visual illusions in clinical settings depends on methodological refinement that has not happened yet [38].

You can hold both halves of this at once. The effect is real enough to be worth studying and nowhere near clean enough to do anything with.

So the honest summary is narrower than the headline. There is a real group level difference in how susceptible people are to a particular class of illusion, it fits what a prediction based account of perception would expect, and it is not a test of anything about any individual person. Nobody should read a paragraph about a mask and conclude something about themselves or anyone they know.

The same caution applies with even more force to autism. A widely cited 2014 proposal argued that autistic perception involves prediction error being weighted too precisely, so that ordinary noise gets treated as meaningful and the world becomes harder to model [42]. It is an elegant idea, and later work extended it into a broader account spanning autism and schizophrenia [43].

Then somebody tested it. A study using a mismatch negativity task with rhythms of varying complexity compared nineteen neurotypical participants against twenty one autistic participants, aged six to twenty one, and looked for the reduced prediction error response the theory forecast. Both groups showed the same stepwise decrease in the response as the rhythms became less predictable. The paper's title states the result plainly: individuals with autism had no detectable deficit in neural markers of prediction error [44]. Later work testing cue-outcome learning in autistic children did not find the clean deficit either [45].

One null result does not sink a hypothesis. But it is exactly the kind of finding that gets left out of the confident version of this story, and the confident version is the one that circulates.

The One Prediction Error Nobody Argues About

Amid all the contested evidence there is a signal that is genuinely nailed down, and it is worth being clear about it because it anchors everything else.

In 1997, Wolfram Schultz, Peter Dayan and Read Montague described what midbrain dopamine neurons in monkeys actually do. They do not signal reward. They signal the difference between the reward you got and the reward you predicted [46]. An unexpected reward drives a burst. A fully predicted reward drives nothing much. A predicted reward that fails to arrive drives a dip below baseline at exactly the moment it was due.

That is a prediction error, measured in single neurons, in a specific pathway, with a mathematical form borrowed from reinforcement learning that fits the data closely. It is the best established prediction error in neuroscience, and nobody serious disputes it.

Human work has since shown that these error signals modulate cortical processing and shape learning about which sensory cues predict which outcomes [47] [48], and the framework has been extended to treat choice itself as a form of inference [49]. There is a great deal more to say about how dopamine teaches the brain than belongs here.

What makes the dopamine result so useful as an anchor is that it did not come from asking whether the brain is Bayesian. It came from recording cells while an animal learned, watching the response drift earlier in time from the reward to the cue that predicted it, and realising that the pattern matched an equation that already existed in machine learning. Two fields converged on the same quantity from opposite directions. That is a much stronger kind of evidence than a theory finding what it went looking for.

It is worth being precise about what this does and does not license. The dopamine result proves that the brain computes prediction errors somewhere, in at least one system, for at least one kind of quantity. It does not prove that sensory cortex works the same way. That extra step is the one currently in dispute.

The Brain Predicts Your Body Too

The framework does not stop at the senses pointed outward. There is a version of it aimed inward, at the signals coming from your organs, and it has become one of the more productive corners of the field.

The proposal is that your brain models the state of your body the same way it models a room, generating predictions about heart rate, gut, breath and temperature and treating the difference as error. Feelings, on this account, are not readouts of a bodily state. They are inferences about the causes of bodily signals [50] [51]. The sense of simply being present in your own body has been given the same treatment [52].

The same logic has been pushed to how the body manages its energy budget. Regulation on this view is not a thermostat reacting to a change but a system spending in advance of a demand it expects, which is why your heart rate climbs before you start running rather than after [53]. It also gives a handle on why people differ so much in how well they read their own internal signals. Some people are simply weighting those signals more heavily than others, and that weighting is precision again [54].

Stop and consider how strange that is as a claim about feelings. It says that hunger and anxiety and the sense of being at home in your own body are not messages arriving from your organs. They are the brain's best account of why those messages arrived.

Pain is where this gets most concrete, because pain is obviously not a simple readout of tissue damage. Studies separating what is expected from what is delivered have found distinguishable contributions of stimulus intensity and prediction in the insula [55], and have traced the temporal and spectral signatures of expectation and prediction error during heat and pain [56].

How Far the Idea Reaches

Part of the appeal here is coverage. Once you have the loop, it starts explaining things in domains that had nothing to do with each other.

Binocular rivalry is a good example. Show each eye a different image and perception does not blend them, it alternates, flipping between one and the other every few seconds. Under a prediction based account this is a system that cannot settle on a single best explanation for its input, so it cycles between two competing hypotheses [57].

The same reading works for music. A piece of music is a stream of expectations being set up, delayed, broken and satisfied, and there is a case that musical pleasure is largely about the handling of those predictions [58].

Language gave the framework one of its cleanest targets. The N400, one of the most studied signals in the whole of cognitive electrophysiology, appears when a word is unexpected in context. A computational model has since been built showing how the N400 falls out of a predictive coding architecture [59].

Memory is a natural fit, because remembering is prediction pointed backwards. The same generative model that guesses what is in front of you can generate what was behind you. It is why memory is reconstructive rather than a recording, and why false memories form as readily as they do. One review argues the two systems are not merely similar but share machinery [60]. A more specific claim is that prediction error is the switch: a memory reopens for editing only when what you recall fails to match what then happens, which is why an incomplete reminder destabilises a memory and a complete one leaves it alone [61]. Modelling work has gone further and derived the sense of how much time has passed from the same predictive machinery [62].

Prediction and memory are closer than they look. Both need a model of how the world tends to go, and both fill in what was never recorded.

Sleep supplies a neat test. If the hierarchy is doing the work, then losing consciousness should break the higher levels first while leaving the lower ones running. That is what one study found: hierarchical predictive coding is disrupted during sleep, with the local, early responses surviving and the higher order ones failing [63].

Sleep is useful here because it hands you a natural experiment. Nobody can ethically switch off a person's higher cortex. You can wait until midnight and it happens anyway.

And then there is action. Active inference treats movement as prediction: you predict the sensory consequences of the movement you intend, and the reflex arcs act to make the prediction come true [64] [65]. That idea has since been built out into a general account of behaviour under the name active inference, in which perceiving and acting are the same computation pointed in opposite directions [66].

Notice the shape of this section. Every one of these is a domain where somebody already had a perfectly decent local theory, and the prediction account arrives offering to replace several local theories with a single one. That is genuinely attractive. It is also exactly the sort of offer that should make you want to see the receipts, which is what the second half of this article is about.

That sequence is the hollow mask in five lines. The mask sends up evidence that it is concave, the error is treated as unreliable next to a lifetime of convex faces, and the prediction survives contact with the data. What you see is the surviving prediction.

The Dark Room

Now the objection.

If the brain exists to minimise prediction error, and prediction error is a measure of surprise, then there is an obvious way to win. Find a dark, silent, featureless room. Go in. Stay there. Nothing happens, nothing is unexpected, prediction error goes to almost zero and stays there for the rest of your life.

Nobody does this. People seek out novelty, travel, difficult music, unfamiliar food and frightening films. This is the dark room problem, and it is the sharpest challenge the framework faces, because it does not attack any particular experiment. It attacks whether the core claim, taken at face value, can possibly be right about behaviour [67].

The standard reply runs through action. In active inference the system is not minimising surprise right now, it is minimising expected surprise over the long run, and a creature that never explores ends up with a model so poor that everything eventually becomes surprising. Starvation is extremely surprising. On this account curiosity is not an exception to prediction error minimisation but a strategy for it.

It is worth taking that answer seriously before deciding whether it works. A creature that only ever minimised surprise moment to moment would stop eating, since hunger is predictable and food is not. It would never explore, and so its model of the world would slowly go out of date while the world kept moving. Eventually the mismatch between the stale model and the actual environment becomes enormous, and the result is not a peaceful dark room but a catastrophe. On that reading, curiosity is prudence.

Critics find this too convenient. One line of argument holds that the dark room only looks like a problem because the framework inherited an old cognitivist picture of an agent separated from its world, and that the problem dissolves if you take the organism and environment as a single coupled system [68]. A related critique argues the framework misdescribes what living systems do, and that the anticipating brain is not running experiments on the world like a scientist [69].

What matters for a reader is less which answer is right and more what the exchange reveals. When a theory can absorb an apparently fatal objection by redefining its central term, you should at least ask what it would take for the theory to lose.

Which is the next question.

What Would Prove It Wrong

In 2023 a group of researchers published a paper in Neuroscience and Biobehavioral Reviews with the title "Is predictive coding falsifiable?" [3]. That is not a rhetorical flourish. It is a real worry about a framework that has become flexible enough to accommodate almost any result.

The concern is easy to state. If a brain region responds more to unexpected stimuli, that is prediction error. If it responds less, that is a fulfilled prediction. If attention changes the sign of the effect, that is precision weighting. If a group shows a smaller effect, their priors are weaker. If they show a larger one, their priors are stronger. Each of those explanations is reasonable on its own. Together they cover the whole space of possible outcomes, and a theory that predicts everything forecasts nothing.

This is not hostile fringe commentary. A 2020 review in the Annals of the New York Academy of Sciences went through the neurophysiological evidence for predictive processing as a model of perception and found it considerably more equivocal than the framework's popularity suggests [70]. A 2017 piece asked whether the brain performs Bayesian inference with predictive coding or without it, and concluded the two are much harder to tell apart than usually assumed [71].

The same pattern shows up when the question is narrowed to one sensory system. A 2018 review of auditory cortex asked directly whether the evidence there supports predictive coding and gave a heavily qualified answer [72]. A 2024 review surveyed the empirical status of predictive coding and active inference across the board and reported a picture that is promising and unfinished rather than established [73].

That is four reviews, in four venues, across four years, arriving at variations of one verdict. Not that the framework is wrong. That the evidence routinely cited for it does not carry the weight being put on it.

The most useful contribution to this argument is a 2021 review that stopped talking in general terms and named the specific confounds. It identified four things that can mimic or inflate the expectation suppression effect: surprise, attention, stimulus repetition and adaptation, and stimulus novelty. Then it went through the experimental literature design by design. Genuine expectation suppression survived in one specific class of study, those where participants learned sequences over weeks before anything was measured. Across the more common designs, where probabilities were learned within one or two sessions, the evidence was inconsistent [74].

Follow that tree honestly and a lot of published findings end up in the orange box. A separate paper asked whether expectation suppression might simply be reduced attention to predictable things, which is a plainer explanation for the same data [75].

Nearly Identical

The single most uncomfortable result for the framework came out of a straightforward experiment.

A team recorded from primary visual cortex and area V4 in macaque monkeys, measuring both spiking activity and local field potentials, and ran a matching study with scalp EEG in humans. They presented a fixed sequence of visual stimuli on most trials and violated the expected order on a few. If sensory cortex contains neurons dedicated to signalling prediction error, a pattern-violating stimulus is exactly the thing that should light them up.

The responses to expected and pattern-violating stimuli were nearly identical. In the spiking, in the field potentials and in the EEG. The authors wrote that their results challenge the assertion that a fundamental computational motif in sensory cortex is to signal prediction errors [4].

Around the same time, another group went after the flagship rodent evidence. Mismatch signals in mouse visual cortex, seen when visual flow coupled to the animal's running unexpectedly stops, had been treated as some of the cleanest prediction error data anywhere. This study showed the same signals could be produced by stimuli that had nothing to do with the animal's own movement, that responses were strongest at each neuron's preferred orientation, and that the whole pattern is explicable by the convergence of known motor and sensory signals. Their conclusion offers a purely sensory and motor explanation for what had been called mismatch signals [76].

The mouse work matters more than it sounds. Rodent recordings are where you get to see single neurons doing the thing, and for years those mismatch signals were the closest anyone had come to catching a prediction error in the act.

Similar deflations have landed elsewhere. When large language models turned out to fit brain data recorded during listening, this was widely read as support for predictive coding, until a 2023 paper argued the fit is better explained by feature discovery than by prediction [77]. In spoken word recognition, a study asked whether the observed reduction in signal actually implies predictive coding and concluded that it does not [78].

There is a pattern worth naming. In each of these cases the original measurement was real and nobody disputes it. What changed was which explanation needed the fewest assumptions, and prediction was not it.

None of this means the framework is wrong. It means that a lot of what has been counted as evidence for it is also consistent with simpler accounts, and that telling them apart is much harder than the popular version admits.

Even the supportive evidence carries hedges. A 2021 meta-analysis pooled the imaging literature and did find a distributed predictive network, which is a genuine result in the framework's favour. In the same abstract the authors report that they found no evidence, at the network level, for a distinction between error processing and prediction processing [79]. That distinction is the entire architecture. Finding the network but not the split is a strange kind of support.

What Is Settled and What Is Not

It would be easy to come away from that section thinking the whole thing collapses. It does not, and pretending otherwise would be the mirror image of the overclaiming this article set out to avoid.

Several things are not in doubt. The brain sends far more signal backwards through cortex than forwards from the senses. Expectation changes the measured neural response to a physically identical stimulus. Dopamine neurons encode reward prediction error. The brain responds to omitted stimuli, to events that did not happen. People with schizophrenia are, at the group level, less susceptible to depth inversion illusions. And perception is not a passive recording, which every serious account agrees on, including the accounts that reject predictive coding as the mechanism.

What is in doubt is narrower and more technical, and it is the part that popular explanations flatten.

There is a useful distinction hiding in that list. Some of it concerns whether prediction happens at all, and that part is not really in question. The rest concerns whether prediction is implemented by the specific mechanism predictive coding describes, with dedicated units passing errors up a hierarchy. You can accept the first without the second. A brain could use expectation heavily and still not do it by subtracting predictions from inputs at every level.

Most of the current argument is about the second thing, which is why it can look, from outside, as though the whole idea is under attack when it is not.

Notice what the right hand column is not. It is not a list of cranks. These are papers in Cerebral Cortex, Cell Reports, Neuropsychology, Autism Research and Neuroscience and Biobehavioral Reviews, mostly written by people working inside the framework rather than against it. That is what a healthy field arguing with itself looks like.

Why the Idea Survives Anyway

So why does predictive processing keep growing if the evidence for its central mechanism is this unresolved?

Because it does something no rival account does. It takes perception, attention, action, learning, emotion, memory and several psychiatric conditions and describes them with one set of terms. Before it, those were separate literatures with separate vocabularies. Whether or not cortex literally implements predict-and-subtract, the framing has generated a decade of experiments that would not otherwise have been run, including the experiments now being used against it. A theory that provokes its own falsification tests is doing more work than one that sits quietly being true.

It also survives because the basic observation underneath it will not go away. Your senses do not deliver enough information to build what you experience. The gap has to be filled from somewhere, and the only available source is the accumulated statistics of everything that has happened to you so far. That much is not really disputable. The argument is about the machinery.

And it survives because of things like the mask.

Go back to it one more time. A rotating hollow face, a laboratory, a person watching it turn. Almost everyone in that room sees a nose pointing at them that is not there. The plastic is telling them the truth and they cannot hear it over their own expectations. Occasionally somebody sits down and reports the shape correctly, and the reason they can is not that they are looking more carefully. It is that the voice supplying the convex face has gone quiet in them.

Whatever the final account of the wiring turns out to be, that is the fact the wiring has to explain. You are not watching the world. You are watching your best guess about it, corrected as it goes, and the correction is usually so fast and so complete that you never notice the guess was there.

There is one more reason to care about how this argument resolves, and it is not academic. If perception really is inference, then disagreements about what happened are not always dishonesty or inattention. Two people standing in the same room with different priors are running different models over the same evidence, and they will sometimes arrive, honestly, at different worlds. That does not make all accounts equally good. It does mean that the confidence you feel about what you saw is generated by the same process that produces the illusion, and it comes with no built in warning when the model has gone wrong.

The mask is the cleanest demonstration of that anyone has built. Everything about the experience says convex. Nothing about the object agrees. And the feeling of certainty is exactly as strong either way.

Most of the time it is right. That is why it feels like seeing.