201 – You have never seen reality

There is an experiment so simple that it requires nothing beyond the page you are currently reading. Focus your eyes on a single word somewhere near the center of this line. Without moving your eyes, try to make out the words at the far left and right margins. They are within your visual field — the light from them is reaching your retina — but they are blurred, uncertain, recognizable at most as shapes rather than legible text. This is not a defect. It is how human vision works. The eye achieves sharp resolution only in a small central region called the fovea, and the impression of uniform clarity across the entire visual field is something the brain constructs by moving the eyes rapidly across the scene and stitching the results together into a seamless whole.¹ What we experience as vision is not a recording. It is an edited summary.

This is a small example of something large. The intuitive model of perception — the sense that we look at the world and see it — is wrong in a specific and consequential way. We do not see the world and then construct a model of it. We run a model of it continuously, and what we call seeing is the output of that model. The distinction sounds philosophical until you encounter a static image in which the brain generates vivid, irresistible movement where none exists. Akiyoshi Kitaoka, a professor of psychology at Ritsumeikan University in Kyoto, has spent decades creating exactly these images. His Rotating Snakes illusion — freely available at www.ritsumei.ac.jp/~akitaoka — consists of a pattern of concentric circles whose carefully arranged luminance gradients cause the brain’s motion-detection system to register rotation in what is a perfectly static image. Before continuing, it is worth taking thirty seconds to look at it. The philosophical claim this article is making about the brain’s generativity will be more viscerally real after that experience than any description can make it.² What makes this illusion more instructive than the standard length-distortion examples is precisely what it shows: the brain is not misreading a signal. It is generating an event that the image itself does not contain. The model is not a passive receiver. It is a creative act.

This article is about what follows from this — about what kind of beings we are, perceptually, and what it means to navigate a world that we can never access directly but only through the particular lens that evolution, development, and experience have built for us. The answer, when examined carefully, is both more interesting and more workable than the standard response to this observation, which is a kind of vertiginous despair. We cannot see reality directly. We can, with practice, see our model of it more clearly. That is not a consolation prize. It is the only game available, and it turns out to be enough.

The prediction machine

The most influential account of how perception actually works comes from the neuroscientist Karl Friston, building on ideas that go back to the nineteenth-century physicist and physiologist Hermann von Helmholtz.³ The core proposal is this: the brain does not primarily receive information and process it into perception. It primarily generates predictions about what incoming sensory data should look like, given its current model of the world, and then uses the actual incoming data to update those predictions. What we consciously experience is not the raw sensory signal but the brain’s best current guess about the state of the world, continuously refined by the mismatch between what was predicted and what actually arrived.

This framework — called predictive processing, or the predictive coding hypothesis — makes sense of a great deal that the simpler picture cannot explain. Consider what happens when you hear your name spoken across a noisy room. The acoustic signal is degraded, partial, competing with dozens of other signals at the same frequency range. A simple recording device would not reliably detect it. Your brain does, reliably, because it is not waiting passively for a signal to cross a threshold — it is actively generating a model of the auditory environment and flagging the specific deviation from prediction that your name represents. Or consider how you manage to read this sentence even though the letters are not perfectly formed, the spacing is not perfectly uniform, and the page is not uniformly lit. The brain fills in what is expected based on what has been seen before, and the result is seamless.

The prediction machine is extraordinarily efficient. By generating expectations rather than processing all incoming data from scratch, it allows a very large amount of complex behavior to proceed with very little conscious attention — the skilled driver who navigates familiar roads while thinking about something else entirely, the musician whose fingers find the notes without conscious direction, the conversationalist who tracks social dynamics that are too rapid and too subtle for deliberate analysis. These are all performances of prediction: the brain running its models ahead of events, allowing the organism to act before the slow machinery of conscious reasoning has had time to catch up.

But the efficiency comes at a cost that is the subject of this article. A prediction machine is, necessarily, a machine that sees what it expects rather than what is there. The errors it makes are not random. They are systematic, patterned, and shaped by everything the model already contains — by prior experience, current expectations, emotional state, and the cultural and social frameworks within which perception occurs. Two people observing the same event with different models see, in a specific and non-trivial sense, different events. This is not a figure of speech. It is a description of the machinery.

What the model contains — and what it omits

The brain’s predictive model is built from everything that has been experienced before: the statistical regularities of the sensory environment, the correlations between actions and their consequences, the social and emotional associations that experience has attached to particular kinds of objects, people, and situations. The model is, in this sense, a compressed representation of the organism’s entire history of interaction with the world.

This means that the model has real content. It is not a blank medium through which reality passes untransformed. It contains assumptions about what kinds of things exist, what they tend to do, what contexts they tend to appear in, and what they tend to mean. A chess grandmaster looking at a mid-game board sees not thirty-two pieces in a configuration but a small number of familiar patterns — the kind of position that tends to lead to a kingside attack, the kind of pawn structure that tends to produce an endgame advantage, the kind of weakness that an opponent is likely to try to exploit. The board the grandmaster sees is not the same board the novice sees, even if the physical arrangement is identical. The grandmaster’s model is richer, more organized, and more precisely calibrated to the relevant features of this particular domain. This is what expertise is, in cognitive terms: not more knowledge stored in the same container, but a more sophisticated model that perceives what the novice’s model cannot.⁴

But the model also has systematic omissions. It is biased toward the expected and tends to miss the unexpected — not because the unexpected fails to arrive at the sensory surface, but because the prediction machine assigns low weight to signals that deviate sharply from what was predicted. This is why eyewitness testimony is systematically unreliable in ways that are not random: people do not remember events neutrally, they remember events as their model interpreted them at the time of encoding, and the model’s interpretation is shaped by expectation, attention, and emotional salience in ways that are invisible to the witness.⁵ It is why we tend to notice evidence that confirms what we already believe and fail to notice evidence that contradicts it — not as a matter of deliberate self-deception, but as a direct consequence of how prediction-based perception works. The prediction machine flags deviations from expectation; it suppresses signals that match it. Information that confirms the model passes through without generating a conscious signal. Information that disconfirms it generates an update — but only if the discrepancy is large enough to overcome the inertia of the existing model.

And here lies the crux of what this series is about. The model that each of us carries is partial — it captures some features of the world and misses others, it foregrounds what prior experience has made salient and backgrounds what it has not, it is calibrated to some domains and uncalibrated to others. The partiality is not a malfunction. It is the necessary consequence of being a finite organism building a workable representation of an environment that contains vastly more information than any nervous system could process. The map is always smaller than the territory. The question is not how to stop this from happening — it cannot be stopped — but whether we are aware that it is happening.

The independence of perception and expectation — and their entanglement

There is an important complication that the straightforward prediction-machine account can obscure. The relationship between what we expect and what we see is not simply one of imposition: the model does not simply override the data entirely. If it did, perception would be indistinguishable from hallucination, and the organism would be unable to learn from experience at all. What actually happens is more subtle: the weight given to incoming data versus the weight given to prior expectation varies continuously depending on the precision or reliability of each source.⁶

In familiar environments, where prior experience is a reliable guide to what is present, the model runs on a longer leash and the incoming data is down-weighted. In novel environments, where prior experience is not a reliable guide, the incoming data is given more weight and the model is more provisional. The skilled navigator in familiar waters trusts the chart. The navigator in uncharted territory trusts the instruments. Both are rational responses to the reliability of available information. The problem arises when the distinction between familiar and novel is itself misclassified — when the environment has changed in ways that the model has not registered, so that the model continues to run with high confidence in conditions where it is no longer reliable.

This is one of the most consequential failure modes of predictive perception, and it generalizes far beyond the purely sensory. The experienced manager whose model of organizational dynamics was built in a previous era applies it to a new organization that operates by different norms and misreads signals that do not fit the old template. The political analyst whose model of the electorate was built on data from previous decades applies it to an electorate that has been transformed by forces the model does not include. The scientist whose model of a field was built during their doctoral training applies it to findings that were produced by methods and instruments that postdate their training. In each case, the model is running in conditions where it is no longer reliable, and the mismatch between its predictions and what is actually present is being suppressed rather than registered. The model cannot see its own obsolescence.

This is not a rare pathology. It is the normal condition of any sufficiently complex model operating in a sufficiently complex environment. The world changes. Models change more slowly. The gap between them is where perception misleads us most systematically, and most invisibly.

What we see when we look at other people — and at ourselves

The predictive model operates not only in the physical and perceptual domain but in the social domain with equal force and considerably less transparency. The models we carry of the people around us — of their intentions, their characters, their likely responses to particular situations — are built from prior experience, cultural templates, emotional associations, and the stories we tell ourselves about the kind of person we are dealing with. These models drive social perception as powerfully as the brain’s visual model drives what we see on a page.

The psychologist Fritz Heider demonstrated in a series of simple experiments in the 1940s that human beings attribute intentions, emotions, and relationships to shapes moving on a screen in ways that the shapes’ physical properties cannot justify.⁷ A large triangle that chases a small circle is not chasing it — there is no intentionality in the relationship between two geometric forms — but the perception of pursuit, threat, and eventual escape is immediate and irresistible. The social perception machinery is not an optional add-on to a neutral physical perception system. It is a primary mode of engagement with the world that operates before and below the level of deliberate reasoning.

The consequence is that other people are never perceived neutrally. They are perceived through a model that has been built from every prior encounter with people, from the cultural templates that structure the social world, from the emotional history that has attached particular associations to particular kinds of person, and from the immediate context that primes certain interpretations over others. The student who expects the teacher to be critical perceives criticism in neutral feedback. The negotiator who expects the counterpart to be adversarial perceives aggression in cooperative overtures. The voter who expects politicians to be corrupt perceives corruption in legitimate political processes. These are not failures of rationality in any simple sense. They are the outputs of prediction machines doing exactly what prediction machines do.

And then there is the most difficult case: the model of the self. Of all the models we run, the one that is furthest from neutral inspection and closest to our deepest commitments is the model of who we are — our character, our capacities, our motivations, our history. This model is the lens through which we interpret our own experience, the framework that makes our choices feel coherent, the narrative that connects the person we were with the person we are and the person we intend to become.

It is also, in a specific way that is easy to overlook, almost entirely untested. Most people’s self-model of their courage was built in conditions that required none. The person who is confident they would help a stranger in danger has probably never been in a situation that carried real personal risk. The person who believes they would stand up to authority has not been through a Milgram experiment. The person who imagines they would remain calm in a crisis has not been in one. And the research literature on this gap — between the self we imagine and the self that emerges under pressure — is consistent and uncomfortable. In a series of studies, Nicholas Epley and David Dunning asked people to predict their own future generous behavior and the behavior of their peers. People systematically overestimated their own generosity and accurately predicted their peers’. The flattering self-model is not an occasional error. It is the default.⁸

This does not mean the self-model is worthless. It means it should be held with the specific calibration that its evidence base warrants: well-tested in familiar conditions, largely extrapolated in unfamiliar ones, and genuinely uncertain about what would happen if the conditions became extreme. The person who discovers in a crisis that they are braver than expected has learned something real. The person who discovers they are more frightened, more selfish, or more compliant than their self-model predicted has also learned something real — and the second discovery is considerably more common than the first, and considerably more useful if registered honestly rather than explained away. The most honest position is not the confident self-model of the hero who acts without hesitation, nor the defeatist self-model of someone who assumes they would fail. It is the genuinely uncertain self-model of a person who has not yet been fully tested and knows it. Article 906 of this series examines the model of the self in more detail.

The impossibility of the view from nowhere — and why it doesn’t matter

There is a philosophical tradition that treats the partiality of human perception as a tragedy: an epistemological fall from a grace that we never actually possessed but somehow feel we should have. On this view, the fact that we see through a model rather than directly is a defect, an obstacle between us and a reality we could in principle access if our instruments were better.

This view is mistaken, and not only philosophically. Neurologically, a brain that processed all incoming sensory information without the guidance of prior expectation would be paralyzed — overwhelmed by the sheer quantity and variety of signals, unable to assign them significance or organize them into actionable perceptions. The model is not an obstacle to seeing the world. It is the condition of seeing it at all. A brain with no prior model would not perceive a pure and unmediated reality. It would perceive chaos.

What this means in practice is that the question is not how to achieve the view from nowhere — the perception unconditioned by prior experience, cultural context, or emotional history. That view is not available to creatures of our kind. The question is how to hold our models with increasing honesty: to know, with as much precision as possible, what they contain and what they omit, where they are well-calibrated and where they are not, and what kind of experience would give us genuine reason to revise them.

This is the practice the series calls The Conscious Look. It does not require escaping the model. It requires turning attention, periodically and deliberately, back onto the model itself — asking what it is predicting, where those predictions come from, and what kinds of signals it would need to encounter to generate a genuine update. It is uncomfortable in the specific way that any honest self-examination is uncomfortable. And it is the only available response to being the kinds of creatures we are: prediction machines whose predictions are always partial, always shaped by what has come before, and always capable of being revised — provided we are willing to look.

Notes

The technical term for this property of the eye is foveal vision. The fovea is a small central region of the retina with a very high density of cone photoreceptors, responsible for detailed color vision. Outside the fovea, visual acuity drops sharply, and color sensitivity is reduced. The brain compensates through saccades — rapid eye movements that bring different parts of the scene into foveal focus three to four times per second — and through a process of spatial integration that stitches these successive fixations into the impression of a uniformly detailed visual field. This impression is a construction; it does not correspond to any single state of the retina.
The Rotating Snakes illusion was created by Akiyoshi Kitaoka, Professor of Psychology at Ritsumeikan University in Kyoto, and is freely available at www.ritsumei.ac.jp/~akitaoka. The image consists of concentric circles whose segments are arranged in a specific asymmetric sequence of luminances — moving from dark to light in one direction around the circle — that causes the brain’s motion-detection system to register rotation in a perfectly static image. The effect is most pronounced in peripheral vision and largely disappears when the eye is held perfectly still on a single point; it returns immediately when the eye moves, because the small involuntary eye movements called saccades interact with the luminance asymmetry to trigger the motion signal anew. The mechanism is not fully understood, but the most widely supported account holds that different luminance levels are processed with different latencies in the visual system, and that this differential latency, when combined with the asymmetric arrangement of the image, is interpreted by the motion-detection system as directional movement. Kitaoka notes on his website that some observers experience dizziness from prolonged viewing — itself an instructive datum, since it shows that the conflict between the perceived motion and the vestibular system’s correct report of no movement is registered not merely cognitively but physically. For a more detailed technical account, see Bach, M. and Atala-Gerard, L. (2020). The rotating snakes illusion is a straightforward consequence of non-linearity in arrays of standard motion detectors. i-Perception, 11(5).
Karl Friston’s free energy principle, which provides the mathematical framework for predictive processing, is one of the most ambitious and technically demanding theories in contemporary neuroscience. Friston argues that biological systems minimize a quantity called free energy, which is related to the mismatch between predicted and actual sensory inputs. The theory is controversial in its strong forms and highly influential in its weaker forms. The most accessible general discussion is Clark, A. (2016). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press. Hermann von Helmholtz’s original proposal that perception is a form of unconscious inference appears in his Treatise on Physiological Optics, volume 3 (1867).
The classic experimental demonstration of expert pattern perception is the research of Adriaan de Groot and later William Chase and Herbert Simon on chess perception. When chess grandmasters and novices are shown a mid-game board position for a few seconds and asked to reconstruct it from memory, grandmasters dramatically outperform novices — but only when the position is a legal game position. When the pieces are placed randomly, the grandmaster advantage disappears entirely. This demonstrates that the grandmaster’s superior memory is not a general memory advantage but a consequence of perceiving the board in terms of meaningful patterns rather than individual pieces. Chase, W. G., and Simon, H. A. (1973). Perception in chess. Cognitive Psychology, 4(1), 55-81.
The unreliability of eyewitness testimony has been demonstrated in an extensive research program by Elizabeth Loftus, beginning in the early 1970s. Loftus showed that the memories of eyewitnesses are not fixed recordings but are continuously reconstructed in the light of subsequent information, suggestion, and expectation. Her research has had significant practical consequences for the criminal justice system, where eyewitness testimony has been shown to be a major contributor to wrongful convictions. For an accessible account, see Loftus, E. (1996). Eyewitness Testimony. Harvard University Press.
The technical term for this weighting process in the predictive processing framework is precision-weighting. The brain does not treat all incoming sensory signals as equally reliable, nor does it treat all predictions as equally certain. It assigns a weight to each that reflects its estimated reliability, and the combined percept is a precision-weighted combination of prediction and data. When sensory data is highly reliable — bright light, clear acoustic signal, familiar environment — the data receives high weight and the prediction is rapidly updated. When sensory data is unreliable — dim light, noisy environment, unfamiliar context — the prior prediction receives higher weight and the data updates it more slowly. This is why we are more susceptible to motion illusions like the Rotating Snakes in peripheral vision, where sensory precision is lower, and why the prediction machine is more resistant to updating in uncertain environments.
The Heider-Simmel experiment (1944) is one of the most cited in social psychology. Fritz Heider and Marianne Simmel showed participants a short film of geometric shapes — a large triangle, a small triangle, and a circle — moving around a rectangle. Participants were asked to describe what they saw. Almost all gave accounts involving intentional agents pursuing goals: the large triangle was chasing the small triangle and the circle, the circle was trying to escape, the small shapes were hiding. The shapes have no intentions; the perception of intentionality is entirely a product of the viewer’s social perception machinery. Heider, F., and Simmel, M. (1944). An experimental study of apparent behavior. American Journal of Psychology, 57(2), 243-259.
The Epley and Dunning studies are reported in Epley, N., and Dunning, D. (2000). Clued in by their peers: People identify positive qualities in others better than they identify them in themselves. Journal of Personality and Social Psychology, 78(3), 497-511. The broader literature on the gap between self-predicted and actual behavior under pressure includes Stanley Milgram’s obedience studies (1963-1974), in which the large majority of participants administered what they believed to be dangerous electric shocks to a stranger when instructed to do so by an authority figure — a result that directly contradicts the self-models of virtually every person tested, almost none of whom predicted they would comply. Jordan Peterson’s observation — that you can only find out what you actually believe by watching how you act — captures the same point from the perspective of clinical psychology rather than experimental method. The self-model is a hypothesis that the situation has not yet tested.