Optical Illusions: What Your Brain's Visual Mistakes Reveal About Normal Seeing

Introduction

In 2022 a vision scientist published a paper with an awkward question in the title. When is an illusion not an illusion?

His answer was blunter than the title. Brian Rogers argued that there is no coherent and meaningful definition of the word at all, and that most of the things filed under it belong in one of three other boxes: effects that are not illusory under any definition, effects that are simply consequences of how our perceptual systems work, and effects that only appear because somebody built an artificial or impoverished stimulus [1].

Read that again. The claim is not that some illusions are misnamed. It is that the category is doing no work.

That is a difficult thing to say in public, because everybody already knows what an optical illusion is. Two lines that are the same length look different. A grid seems to shimmer. A dress splits the internet. The standard explanation follows immediately: your brain got confused, your eyes played a trick, the visual system glitched.

Almost none of that survives contact with the literature.

It is worth being clear about what is being disputed and what is not. Nobody argues that the two lines are actually different lengths. Nobody argues that you are imagining the effect. The measurements are not in question. What is in question is the story we attach to them, and that story has been repeated so consistently for so long that it now sounds like a description rather than an interpretation.

Here is a result from the other end of the field, and it is the cleanest support for Rogers you will find anywhere. In 2026 researchers took a deep convolutional neural network and trained it for one job: lightness constancy. Work out how bright a surface actually is, regardless of how it happens to be lit. That is the useful thing a visual system needs to do. Nobody trained it on illusions, nobody built illusions into it, nobody wanted illusions. The network became susceptible to lightness illusions anyway [2].

Nothing in that network is confused. It has no eyes to play tricks with and no brain to glitch. It learned to solve a hard problem well, and the illusion fell out of the solution.

That is the argument of this article. An illusion is not evidence of a broken visual system. It is evidence of a working one, caught in the act.

The rest of this is what that actually means, where it holds, where it is still being fought over, and what happens to the popular version of the story once you look at the studies underneath it. Some of the most famous claims about illusions turn out to be wrong. One of them is in the textbook you probably learned from.

The Word That Does Not Survive Inspection

Start with the vocabulary, because it does more damage than people realise.

Call something an illusion and you have already made a claim. You have said there is a correct answer, that perception missed it, and that the miss is the interesting part. Every one of those three moves is doing hidden work.

Take the correct answer first. When two lines of equal physical length look unequal, the ruler is treated as the truth and your experience as the error. But the ruler is measuring a flat printed page. Your visual system is not built to report the properties of flat printed pages. It is built to report the properties of the three-dimensional world that flat printed pages are standing in for, and by that standard it is not obvious who is wrong.

Daniele Zavagno and colleagues made this argument carefully across two papers, treating illusions as a problem in how we conceptualise perceptual experience rather than as a catalogue of failures [3]. Zavagno later reframed the whole family as a cognitive clash rooted in perception, which puts the conflict between what you see and what you know, not inside the seeing itself [4]. Philosophers have had the same trouble. Fiona Macpherson and Clare Batty went back to the definitions of illusion and hallucination and found that new cases keep breaking them, which is what usually happens when a category was drawn around examples rather than around a mechanism [5].

None of this is word games. The framing decides what you go looking for.

If illusions are failures, you study them to find the bug. If illusions are the normal output of a system meeting an unusual input, you study them to find out how the system works. The second framing is the one that has actually produced results, and it is the one almost no popular explainer uses.

The classical position on the other side belongs to Richard Gregory, who spent decades arguing that illusions happen when knowledge gets applied where it does not belong [6]. His version keeps the word and keeps the sense of error. It is worth knowing because it dominated the field for a generation, and because the specific case he built it on has since come apart. That story is coming.

That decision tree is Rogers's argument drawn out. Follow any famous illusion down it and watch how rarely you land in the red box.

What Actually Reaches Your Brain

To understand why the red box stays empty, you have to be honest about the input.

Light lands on your retina. That is it. A two-dimensional pattern of intensities and wavelengths, changing constantly, smeared by eye movement, interrupted several times a second by blinks, and missing a chunk where the optic nerve leaves.

From that, you get a stable world of solid objects at known distances.

The gap between those two sentences is the entire problem, and it has a name. The retinal image is ambiguous. An infinite number of physical arrangements could have produced any given pattern of light. A small nearby object and a large distant one can land identically. A dark surface in bright light and a bright surface in shadow can send the same amount of light to your eye. Nothing in the image itself settles which one is out there.

So the visual system cannot read the image. It has to bet on what caused it.

This is not a modern idea. Hermann von Helmholtz called it unconscious inference in the nineteenth century, and the phrase has aged well because it says exactly the right thing: something inference-shaped is happening, and you have no access to it. He had no way to test it. The machinery to do that arrived a century later, and the anatomy it revealed is covered in more detail in our piece on the visual cortex and its two processing streams.

Speed makes the problem worse. The system does not have the luxury of deliberating, because the whole point is to act on the world while the world is still there. Recognition of a scene's gist happens fast enough that any careful weighing of alternatives is off the table, which we cover separately in our article on how quickly the brain recognises a visual pattern.

Fast and ambiguous. That combination has exactly one solution. Guess well.

Notice what that diagram does not contain. There is no step where the system checks its answer against reality. There is no error signal from the world saying you got this one wrong. The output of the guess is not a proposal you get to review. It is simply what you see.

That is why an illusion feels like seeing rather than like being mistaken. You are not shown the evidence and the conclusion. You are shown the conclusion.

You can test the scale of this on yourself in about ten seconds. Close one eye, hold your gaze steady, and think about the blind spot sitting a little off centre in the eye that is still open. There is a patch of your visual field with no receptors in it at all. You do not experience a hole there. You do not experience a grey smudge or a blur. You experience wallpaper, or sky, or whatever the surrounding region happens to be, seamlessly continued across a gap that carries no information whatsoever. Nobody finds this alarming because nobody notices it.

The Guess Machine

The modern name for this is predictive processing, and it is worth being careful here, because it is the current dominant account rather than a settled fact.

The core claim is that perception runs top-down as much as bottom-up. Higher levels of the visual system generate expectations about what the lower levels should be receiving, the lower levels report the difference between expectation and input, and what you experience is the settled result. Matthew and Joseph Nour laid out how illusions fit into this Bayesian framing, where prior beliefs and incoming evidence are combined in proportion to how reliable each one is [7].

The reliability weighting is the part that matters for illusions. When sensory evidence is weak or ambiguous, the prior gets more say. When evidence is crisp, the prior gets less. An illusion is usually a stimulus that makes the evidence ambiguous in a very specific way and then lets a strong prior fill the gap.

There is direct evidence that visual cortex carries this kind of uncertainty rather than just carrying an image. Ruben van Bergen and Janneke Jehee found that activity in human visual cortex reflects the uncertainty of the decision being made, not only the stimulus being viewed [8].

That distinction is easy to skim past. A picture of a stimulus in the brain would be a copy. A representation of uncertainty is a bet with a confidence attached to it, and the confidence is what decides how much the incoming evidence gets to argue back.

The past leaks into the present too. Mauro Manassi and David Whitney describe continuity fields, in which what you saw a moment ago biases what you see now, producing a positive serial dependence [9]. That sounds like a defect. It is closer to a feature, because the real world genuinely is continuous, and a system that smooths across moments will be right more often than one that treats every instant as new. Perception does this outside vision as well, which is a useful sanity check on the whole framework. Rui Zhe Goh, Ian Phillips and Chaz Firestone ran silence through illusions originally designed for sounds, and the silences behaved like sounds do [10]. You do not merely notice that a sound stopped. Something in you represents the gap.

If that all sounds abstract, the next section is the opposite. It is one specific illusion, drawn by one specific man in 1889, and the argument over what it means has been running for more than a century without resolving.

The Most Famous Illusion Has No Agreed Explanation

You already know the figure. Two horizontal lines of identical length, one with arrowheads pointing out at each end, one with arrow tails pointing in. The line with the tails looks longer. It looks longer even when the arrows are removed and replaced with something else, and it looks longer to almost everyone.

This is the Müller-Lyer illusion, and it is in every introductory textbook.

Here is what the textbooks usually do not tell you. The paper that offered the most influential modern account of it opens by calling it the best known and most controversial of the classical geometrical illusions [11]. Most controversial. Not most solved.

The explanation you were probably taught is Gregory's. The arrowheads act as depth cues. Fins pointing outward look like the near corner of a building coming toward you, fins pointing inward look like the far corner of a room going away from you, and the visual system applies size constancy scaling, which is the correction that keeps a person the same apparent size as they walk away from you. Apply that correction where no depth actually exists and you get a length error. Misapplied size constancy [6].

It is a satisfying story. It has been under attack since the 1960s.

W. H. N. Hotopf took the size constancy theory apart in the British Journal of Psychology in 1966, and L. B. Brown and L. Houssiadas had already published a competing treatment of illusory perception as a constancy phenomenon in Nature two years before that [12] [13]. The problem was never that constancy is irrelevant. It was that constancy explanations were being written to fit the illusion after the fact.

This is a familiar failure in perception research and it is worth naming. An explanation that can be adjusted to fit whatever the illusion does is not really an explanation. It has to predict something it was not built from.

Then there is the touch problem. Ray Over showed haptic illusions behaving like inappropriate constancy scaling in 1967 [14], and the Müller-Lyer effect survives being felt rather than seen. Alberto Gallace and Charles Spence traced its crossmodal consequences and found the effect crossing between vision and touch [15]. An explanation built entirely on visual depth cues has to work quite hard to account for an illusion your fingers also fall for. Ross Day proposed a different mechanism altogether, in which the fins contribute to the perceived extent of the whole figure and the apparent length follows the overall shape. Day and Hannelore Knuth also went back to what Müller-Lyer himself actually claimed, which turns out to differ from what he is usually credited with [16].

And then in 2005 the argument changed shape entirely.

Catherine Howe and Dale Purves did not propose a new mechanism inside the head. They went out and measured the world. Using a database of natural scenes with real distance information attached to every point, they asked a statistical question: given a retinal image containing this pattern of lines and fins, what are the physical arrangements that could have produced it, and how long are those arrangements typically? The perceptual effects of the Müller-Lyer figure and its major variants fell straight out of those probability distributions [11].

Their conclusion is the sentence this whole article is built on. The illusion is a manifestation of the probabilistic strategy of visual processing that evolved to deal with the uncertain provenance of retinal stimuli.

Read the implication carefully. On this account, you are not making an error. You are reporting the most likely length, given everything that pattern of light has historically meant. The line really is usually longer when it comes with those fins attached. It just is not longer on the page.

One detail matters for how you weigh this. Howe and Purves did not run human participants for that result. Their sample was a scene database, and the human perceptual data they predicted came from the existing literature. It is a strong result of a particular kind, and it is not a psychophysics experiment.

The dispute is not settled. Bruno and Bernardis published a paper whose title asks why the data on this illusion are so contradictory, which is the honest summary of where the behavioural literature sits [17]. Brain imaging has mapped the cortical activation that accompanies the effect without adjudicating between the accounts [18].

And in 2025 the argument acquired a fresh front. Dorsa Amir and Chaz Firestone asked in Psychological Review whether visual perception is WEIRD, taking direct aim at the cultural by-product hypothesis for this illusion [19]. More on that later, because it deserves its own section.

Notice what that timeline is not. It is not a march toward the truth. It is a question that keeps getting reopened, by better methods each time.

Gregory himself was not only a theorist, and his empirical work has held up better than his general theory. With Priscilla Heard he explained the cafe wall illusion, where straight mortar lines between offset rows of tiles appear to slope, through border locking at the boundaries between light and dark tiles [20]. That account is mechanical, local and testable. It is a good reminder that the field's disagreement is about which explanations generalise, not about whether careful work was done.

When The Argument Moved To The Hand

By the mid 1990s a different question had taken over, and it is the sharpest test anyone has devised for what an illusion actually is.

Forget what the illusion looks like. Ask what your hand does.

If you reach out to pick up the central circle of an Ebbinghaus figure, does your grip aperture match the circle's real size or its illusory size? The stakes are high. If perception is fooled and action is not, then vision is not one system reporting one answer. It is at least two, and only one of them is what you experience.

Angela Haffenden, Karen Schiff and Melvyn Goodale reported exactly that dissociation in the Ebbinghaus illusion in 2001 [21]. Goodale and David Westwood built it into a broader account of duplex vision, in which separate but interacting cortical pathways handle perception and action [22]. David Milner and Richard Dyde asked the natural follow-up: why do some illusions affect visually guided action while others do not [23].

It was a beautiful result. It was also contested almost immediately.

Volker Franz, Karl Gegenfurtner, H. H. Bülthoff and M. Fahle published a paper in 2000 titled, without much diplomacy, "Grasping Visual Illusions: No Evidence for a Dissociation Between Perception and Action" [24]. Their argument was methodological. The perceptual task and the grasping task were not being compared on equal terms, and once you match them properly the difference shrinks or disappears.

That kind of objection is unglamorous and it is usually right. Most of the time in psychophysics the fight is not about what the brain does. It is about whether the two tasks you compared were asking the same question.

Eight years later Franz and Gegenfurtner came back with a paper subtitled "Consistent data and no dissociation" [25]. In between, the specific escape routes were tested and closed one by one. One suggestion had been that grasping only looked immune because the surrounding circles were being treated as obstacles to avoid rather than as context, which would explain the result without any split between perception and action. That was tested directly and ruled out [26].

What is striking about this literature is how careful both camps are. Nobody is being sloppy. The measurements are fine, the participants are real, the effects are reported honestly. The disagreement lives one level up, in what counts as a fair comparison between looking at something and reaching for it, and that turns out to be a genuinely hard question rather than a technicality.

David Carey framed the whole thing as an open question in Trends in Cognitive Sciences, asking simply whether action systems resist visual illusions [27]. Westwood was still asking where we are now in 2008 [28]. Later work found errors in interception could be predicted from errors in perception, which points against a clean separation [29].

Thirty years. Two camps. Good data on both sides. No convergence.

That deserves to be said plainly, because it is the opposite of how illusions are usually presented. This is not a solved topic with a fun demonstration attached. It is a live scientific argument in which the demonstration is the evidence.

Five arguments. Every one of them still running. If you have only ever met optical illusions as a gallery of pictures, that table is the thing worth taking away.

Light, Edges, And The Illusions Built Before You Notice

Move away from geometry for a moment, because the lightness illusions are where the mechanism gets easiest to see.

Two patches print identical grey. Surround one with dark and the other with light and they no longer look the same. This is simultaneous brightness contrast, and it is old, simple and stubborn.

The competing explanations split into families. One says the visual system anchors lightness to something in the scene, typically the brightest surface, and works outward from there. Another says the effect falls out of spatial filtering, the way the system responds to differences across space rather than to absolute amounts. Both accounts have been tested directly against the same data [30].

The two families make the same prediction most of the time, which is why the argument has lasted. Where they come apart is at the awkward stimuli, and that is exactly where somebody had to go looking.

Barbara Blakeslee and Mark McCourt worked this problem for years, analysing brightness induction across space and time [31], and then stepping back to ask what these illusions tell us about the inverse problem of achromatic perception [32].

That phrase, the inverse problem, is the same ambiguity from earlier wearing different clothes. You get light off a surface. You want to know the surface. Illumination and reflectance are multiplied together in the signal and you have to pull them apart with only the product in hand. There is no arithmetic that recovers two numbers from their product, so the system has to bring something else to the table, and what it brings is assumptions about how surfaces and light usually behave.

Colour spreads too. Baingio Pinna, Gavin Brelstaff and Lothar Spillmann described the watercolour illusion, in which a faint coloured contour causes a very large enclosed area to take on a tint that is not physically there [33]. The Craik-O'Brien-Cornsweet effect does something similar with lightness: manipulate a thin edge profile and two large regions that are physically identical look different [34].

Both of those are worth sitting with. The information that changes your experience of a large area is contained in a thin line at its border. The interior is not being measured. It is being inferred from its edge.

Now the experiment that separates the levels.

Julia Harris, D. Samuel Schwarzkopf, Chen Song and Bahador Bahrami used an interocular masking technique that let them make only the context invisible while leaving the target visible [35]. If you cannot consciously see the surround, does the surround still change what you see?

For simultaneous brightness contrast, yes. The illusion persisted with the context masked from awareness. For the Kanizsa triangle, no. When the inducing elements were invisible, the illusory contours did not appear.

Two illusions, one experiment, cleanly separated. One is built early, below consciousness, from local contrast. The other needs the context to reach awareness before the figure can be constructed. Calling both of them "the brain being fooled" hides the most interesting fact about them.

Individual differences confirm the split from another direction. A 2025 factor analysis of lightness illusion magnitudes across two online experiments found the illusions grouping into a few distinct clusters rather than all loading on a single underlying factor [36]. Different illusions, different machinery.

And then the network. Train a deep convolutional network purely to recover lightness constancy from images and it acquires susceptibility to lightness illusions as a side effect [2]. No retina. No cortex. No confusion available to it. The illusion is a property of the solution to the problem, not a property of the hardware solving it.

The Shapes That Are Not There

The Kanizsa triangle is the illusion that best refuses the word illusion.

Three black discs, each with a wedge cut out, arranged so the wedges face inward. You see a white triangle sitting on top of them. Its edges are crisp. Its surface looks slightly brighter than the page. Neither the edges nor the brightness difference exists in the image.

This is not a failure to notice something. It is the active construction of something.

The question of where that construction happens has a good answer. Marianne Maertens and Stefan Pollmann showed that illusory contours do not pass through the blind spot, and used that to support V1 as obligatory for discriminating the curvature of illusory contours [37]. The blind spot is where your optic nerve exits and there are no photoreceptors. If the illusory contour were purely a high-level inference, it would have no reason to care about a hole in the retina. It cares.

Think about how odd that is for a second. The triangle you are seeing does not exist, and yet it respects an anatomical detail of your own eye that you have never noticed and cannot see. Whatever is drawing it is drawing it early.

V1 cannot be doing this alone, though, because a contour that closes into a shape is a global fact and V1 views the world through very small windows. Work on shape processing area LO, which responds to whole objects rather than to edges, tied that region to illusory contour processing as well [38]. So the figure is being assembled at two levels at once, and the real question is how those levels meet.

Stephen Grossberg's answer is that they are complementary rather than sequential, with illusory depth from brightness and apparent motion of illusory contours arising from processes that each supply what the other lacks [39]. Lothar Spillmann and colleagues offered something more concrete, arguing for Gestalt neurons that carry out the local-to-global transformation directly instead of leaving it to be inferred [40].

The most precise recent evidence comes from cortical layers. Johanna Bergmann and colleagues used laminar fMRI across two studies to ask where illusory and imagined content live inside V1 [41]. Illusory content was decodable mainly from the superficial layers. Mental imagery was decodable mainly from the deep layers. Illusory content shared information with real perceptual content. Imagery did not.

That is a striking result and it is worth stating what it means. Imagining a triangle and seeing an illusory triangle feel completely different to you, and they are handled by partially different microcircuits in the very first cortical area that processes vision. The difference in your experience is not a story you tell afterwards. It is built into the wiring.

The dashed arrows are the part that changes everything. Vision is not a one-way trip from eye to understanding. The higher levels talk back, and what you see is the result of that conversation rather than the result of the first message in it.

The same construction happens in damaged eyes, and there it stops being a curiosity. Elisa De Stefani and colleagues found illusory contours forming over pathological retinal scotomas, which are blind patches caused by disease rather than by anatomy [42]. The filling-in is not decorating a hole. It is covering a gap that would otherwise sit in the middle of somebody's vision, and one consequence is that a person can lose a real amount of retina before noticing anything is wrong. Illusory stimuli have since been turned into a mapping tool, using exactly where the fill breaks down to locate the damage [43].

None of that would be possible if early vision were a fixed grid of detectors, and it is not one. Pettet and Charles Gilbert showed that receptive fields in primary visual cortex change size dynamically, so the basic units of early vision reorganise according to what surrounds them [44]. That was work in cats rather than humans, and it should be read as evidence about mammalian visual cortex in general rather than as a human measurement. What it carries is that even the lowest level of the system is contextual. No stage of vision simply reports the pixels.

Motion That Is Not Moving

Stare at a waterfall for a minute and look at the rocks beside it. The rocks drift upward.

This is the motion aftereffect, and it is one of the oldest recorded illusions. Frans Verstraten traced its documented history back much further than most people assume [45]. Ancient observers wrote it down. It has been rediscovered repeatedly ever since, which tells you something about how reliably it works. The standard explanation is adaptation. Motion-sensitive neurons tuned to one direction fire hard during exposure and then respond less afterwards, so the balance between opposing direction detectors tips and a stationary scene reads as moving the other way. Verstraten and colleagues probed the limits of this with high-speed motion and found the aftereffect behaving in ways a simple fatigue account struggles with [46].

Then it gets strange in three separate directions.

Jonathan Winawer, Alexander Huk and Lera Boroditsky produced a motion aftereffect from visual imagery of motion [47]. Participants imagined movement and got the aftereffect. No moving stimulus was ever presented.

Sit with that one. Something you did entirely inside your head adapted a mechanism that is supposed to be responding to light.

Christopher Berger and Henrik Ehrsson found that auditory motion elicits a visual motion aftereffect [48]. Sound moving through space produced a visual aftereffect. And Kunchen Xiao and colleagues found cross-modal motion aftereffects transferring between vision and touch in early deaf adults [49]. Touch and vision, exchanging adaptation.

Whatever is adapting is not a bank of dumb detectors wired to the retina. It is something closer to a representation of motion itself, reachable from imagination, from hearing and from touch.

This has practical edges as well. A 2026 study found that a common prescription lens correction causes motion illusions in presbyopic and general populations [50]. That is not a laboratory curiosity. That is spectacles.

Why Knowing Does Not Switch It Off

Here is the fact that most annoys people, and the one that most clearly separates perception from belief.

You can know an illusion is an illusion. You can have measured the lines yourself, twice, with a ruler you trust. Look again and it is still there.

Philosophers call the general issue cognitive penetrability, and Robert McCauley and Joseph Henrich used the Müller-Lyer as the test case for whether the visual input system is penetrable by what you know, over time as well as in the moment [51]. Their case is that the visual system is substantially insulated from your beliefs, which is why the illusion survives the ruler.

That insulation is not a defect. Consider the alternative. A visual system that updated freely on what you expected to be there would be a system you could talk yourself into seeing things with. Insulation is what keeps perception anchored to evidence.

But the story is not that simple, and this is where the article has to be careful rather than tidy.

Some illusions are modulated by meaning. Nataly Davidson Litvak and Liad Mudrik found that semantic priming affects the probability of experiencing the Kanizsa illusion [52], and a 2025 study found semantic priming modulating both the strength and the direction of the same illusion [53]. Prime somebody appropriately and the illusory figure becomes more or less likely to form.

So the honest position has two halves. Knowing the answer does not dissolve the illusion. Meaning and expectation can nevertheless shift some illusions, and the ones they shift are the ones built later, in the parts of the system where context is being assembled.

Harris and colleagues gave you the reason in the lightness section. Brightness contrast did not need conscious access to its context. The Kanizsa figure did. Illusions built early are sealed off. Illusions built late are not.

That is a much more useful answer than "your brain is stubborn". It also connects directly to a wider pattern, because confident wrongness is not confined to vision. Our article on the illusion of knowing covers the same shape of problem in your sense of what you understand, and the piece on how false memories form shows reconstruction producing confident errors in memory rather than in sight.

Is Illusion Susceptibility Even One Thing?

Ask most people how susceptible they are to optical illusions and they will answer as though it were a single trait, like height or reaction time.

The evidence for that is genuinely mixed, and the mix is interesting.

Lukasz Grzeczkowski and colleagues measured the magnitudes of several visual illusions in the same observers and found no correlations between them [54]. If falling hard for one illusion told you nothing about whether you fall hard for another, then there is no single dial and the popular framing collapses.

Against that, Dominique Makowski and colleagues built a parametric framework that could generate ten classic illusions at varying strengths (Delboeuf, Ebbinghaus, Rod and Frame, Vertical-Horizontal, Zöllner, White, Müller-Lyer, Ponzo, Poggendorff and Contrast) and tested 250 participants inside a perceptual discrimination task [55]. They found evidence for a general factor, which they labelled Factor i, along with personality correlates.

Both studies are careful. They disagree. The 2025 lightness work suggesting several distinct groups rather than one factor sits somewhere between them.

The question is older than any of this, incidentally. Ludwig Immergluck was relating resistance to optical illusion to field dependence in 1966 [56]. People have wanted illusion susceptibility to be a stable personal trait for a very long time.

Two recent findings put real numbers into the debate.

The first is genetic. Lihong Chen and colleagues combined the classic twin method with multichannel functional near-infrared spectroscopy and found that genes account for over half the variance in the strength of the experienced Ebbinghaus illusion [57]. The detail underneath that headline is better than the headline. Activation evoked in early visual cortex was explained by genetic factors, while activation in posterior temporal cortex was explained by environmental ones. Feedforward connectivity from occipital to temporal cortex was modulated by genes. Feedback connectivity was shaped entirely by environment.

Read that last pair again. The signals going up were substantially inherited. The signals coming back down were built by experience.

The second is expertise. Radoslaw Wincza and colleagues tested 44 medical image experts, made up of reporting radiographers, trainee radiologists and certified radiologists, against 107 psychology and medical students [58]. The experts were significantly less susceptible to the Ebbinghaus, Ponzo and Müller-Lyer illusions.

But not to all of them. Susceptibility to the Shepard tabletops illusion was unchanged. The authors suggest the advantage comes from a stronger local processing bias, a habit of ignoring irrelevant context that radiology training builds, and it transfers only to illusions where ignoring context helps.

That is a real finding with a real boundary, and the boundary is the informative part. Nobody became immune to illusions. A specific perceptual habit transferred to a specific subset of them.

Susceptibility also moves across the lifespan. Yarden Mazuz, Yoav Kessler and Tzvi Ganel documented age-related changes in susceptibility to size illusions [59]. Learned visual expertise reshaping perception is a general principle rather than a radiology quirk, and our article on the neuroscience of reading and the visual word form area follows the same principle into literacy.

Attention belongs in this picture too, and it complicates the individual-differences story. Two people looking at the same figure are not necessarily sampling it the same way. Where the eyes go and what gets weighted changes what the inference has to work with, so some of the spread between observers may sit in attention rather than in the visual system itself. Our piece on attention and memory follows that gating in more detail.

The Culture Question That Refuses To Close

One claim about optical illusions gets repeated more than almost any other, and it deserves a careful look because it is genuinely uncertain.

The claim is that people raised in environments full of straight lines and right angles are more susceptible to the Müller-Lyer illusion than people raised in environments without them. The carpentered world hypothesis. If your visual system learned its statistics from rectangular rooms and city blocks, so the story goes, it will read the fins as corners, and if it did not, it will not.

It is an appealing idea, it fits the inference framework beautifully, and the evidence for it is weaker than its fame suggests.

Deregowski spent a career on cross-cultural perception of real and represented space, and his Behavioral and Brain Sciences paper drew the kind of extended peer commentary that only genuinely unsettled questions attract [60]. Marc Bornstein looked at the psychophysiological component of cultural differences in colour naming and illusion susceptibility as early as 1973, and even then the question was whether the differences were cultural, physiological or an artefact of how the tests were run [61].

The methodological problems in the older literature are serious. Participants who have never seen a line drawing are being asked to make judgements about line drawings. Instructions cross languages. Comparison groups differ in schooling, in testing familiarity and in a dozen other things at once. None of that means the effect is not real. It means the older studies cannot settle it.

Two very recent pieces of work reopened the case rather than closing it. Amir and Firestone's 2025 Psychological Review paper asks directly whether visual perception is WEIRD and takes apart the cultural byproduct hypothesis for this illusion [19]. A 2026 paper returns to what the carpentered world does to visual perception with fresh eyes [62]. Developmental work has begun tracking how visual attention to the Ebbinghaus illusion develops across two cultures rather than comparing adults after the fact [63].

So the correct thing to say is that this is open. Not debunked, not established. If somebody tells you confidently that people in non-carpentered environments do not see the Müller-Lyer illusion, they are ahead of the evidence.

When Perception Works Differently

Illusion tasks have been used for decades to probe conditions in which perception is thought to work differently. This is the part of the topic where overclaiming is easiest and most harmful, so it needs care.

Two things are true at once. There are real group-level findings here. None of them tells you anything about any individual, and none of them is a test of anything.

In schizophrenia the literature is large. Ana Luísa Lamounier Costa and colleagues published a systematic review covering 45 studies of visual illusion perception in patients [64]. They report concordant evidence of abnormal processing across most illusion categories, with facial depth inversion and the Müller-Lyer standing out. They also report significant methodological disparities across the studies, and they open by noting that recent conflicting accounts have called the validity of the whole approach into question. That is a review being honest about its own evidence base.

Reviews rarely undercut themselves in the opening paragraph, so when one does, it is worth taking the warning at face value rather than quoting the headline finding and moving on.

Charles-Edouard Notredame and colleagues had earlier argued the case for what visual illusions can teach us about schizophrenia [65]. The predictive coding account behind that work proposes that abnormal prior beliefs drive psychotic symptoms, and illusions are a natural place to test it. The test has not come back cleanly. Mariia Kaliuzhna and colleagues looked specifically for abnormal priors in early vision in schizophrenia and reported finding none [66]. That paper has to be read alongside the review, not instead of it.

The autism literature has moved further, and it has moved against the popular version.

For years the claim circulated that autistic people are less susceptible to visual illusions, which was taken as evidence for reduced use of context or weaker prior expectations in autistic perception. Catherine Manning and colleagues tested this properly, measuring susceptibility to the Ebbinghaus and Müller-Lyer illusions in autistic children aged 6 to 14 and in typically developing children matched on age and non-verbal ability, using three different methods [67]. The point of using three methods was that earlier group differences might reflect differences in decision-making strategy rather than differences in perception. Jon Brock had already raised alternative Bayesian accounts of autistic perception in a commentary on the influential weak-priors proposal [68].

Then in 2025 Yarden Mazuz, Bat-Sheva Hadad and Tzvi Ganel ran a standardised psychophysical battery on 81 participants, 41 autistic and 40 non-autistic, measuring susceptibility to the Ponzo, Ebbinghaus and height-width illusions [69]. Both groups showed clear susceptibility. There was no significant group difference in illusion magnitude, and none in discrimination thresholds either.

The paper is titled "Intact Susceptibility to Visual Illusions in Autistic Individuals". That is a direct challenge to a claim that has been repeated in popular writing for a decade.

The most informative work here runs in the opposite direction, starting from a known lesion and asking what changes. Reduced likelihood of the Poggendorff illusion has been reported in cerebellar strokes alongside imaging of the damage [70]. That is a stronger design than a group comparison, because what differs between the groups is identified rather than inferred.

Parkinson's disease has been approached from both ends, and the contrast is instructive. Susceptibility to geometrical illusions has been measured in the laboratory [71], while separately an interview survey asked patients which visual illusions they actually experience as part of living with the condition [72]. Those two methods answer different questions and neither replaces the other. A task tells you about a mechanism. An interview tells you what somebody is dealing with at home.

Take all of that as a map of where researchers are looking, not as a set of conclusions about anybody. Every finding above is a difference between group averages, several of them are contested, and none of them is remotely close to being usable as an individual measure. If you find yourself seeing an illusion strongly or weakly, that number means nothing about you on its own.

Other Eyes, And Machines That Never Had Any

If illusions were a quirk of human brains, they should stop at the species boundary. They do not.

Oxána Bánszegi and colleagues ran a meta-analysis of responses to geometrical visual illusions across non-human animals and confirmed that illusion perception is a general phenomenon rather than a scattering of one-off reports [73]. Two of their moderator findings are worth having. Studies on birds report stronger illusion perception than studies on other classes, as do studies on animals with laterally placed eyes compared with forward-facing eyes.

One more of their findings deserves quoting for its honesty. Studies with larger samples reported smaller effect sizes. That is the classic signature of publication bias, and the authors say so rather than burying it.

Laura Kelley and J. L. Kelley had earlier argued for taking a proper perceptual perspective on animal illusion and confusion instead of assuming other species see as we do [74]. Since then the specific results have piled up. Dogs are susceptible to the Kanizsa triangle [75]. A citizen science study asked cat owners to lay out illusory square outlines on the floor and found domestic cats sitting in them much as they sit in real squares [76].

Then there are the machines, and this is where the argument closes.

Jinyu Fan and Yi Zeng built an image distortion out of the abutting grating illusion, in which illusory contours emerge from line gratings meeting each other, and used it to test deep learning models [77]. They tested models trained from scratch alongside 109 pretrained models covering various data augmentation strategies. The distortion was challenging even for state of the art systems, and DeepAugment models handled it best.

Put that beside the lightness constancy network from the introduction and the pattern is hard to argue with. Systems with no biology at all, trained on ordinary visual tasks, develop illusion-like behaviour. Not because anyone wanted it. Because the illusions are what you get when you build something that solves the inverse problem using regularities in images.

A brain is one solution to that problem. It is not the only one, and the illusions are not the brain's fault.

Where This Actually Matters

Most illusions cost you nothing. A few do not.

Vebjørn Ekroll and colleagues described what they call the illusion of absence: the compelling and immediate experience that the space behind an occluding object is empty [78]. The region is hidden. You cannot possibly know what is in it. Yet the experience is not one of uncertainty, it is one of having checked.

Their proposal is that this makes blind spots in a driver's field of view more dangerous than the lack of visibility alone would predict, because the driver feels they have already verified the space is clear. It is a well-argued idea and the authors state plainly that the hypothesis requires further testing. Treat it as a proposal, not as a demonstrated cause of crashes.

The reason it is worth taking seriously anyway is that it names a specific psychological state rather than a general risk. Being unsure whether the road is clear and feeling certain that it is clear produce very different behaviour, and only one of them makes you slow down. If the visual system delivers occluded space as confidently empty rather than as unknown, then the dangerous moment is not the one where you cannot see. It is the one where you do not feel the need to look.

Magic has been running on this territory for centuries. Gustav Kuhn and Michael Land made the case that there is more to magic than meets the eye, and that stage magicians have accumulated a working knowledge of perceptual and attentional limits that vision scientists can learn from [79]. Magicians do not fight your visual system. They feed it exactly the input it will misread and then step out of the way. Medicine has its own version. Richard Daffner catalogued visual illusions in the interpretation of the radiographic image back in 1989 [80], describing how contrast and context can make a radiologist see structure that is not there or miss structure that is. Set that next to the finding that experienced radiologists are less susceptible to three classic illusions and a satisfying loop closes. The training that reduces context effects is training against exactly the failure mode Daffner was documenting.

And the spectacles finding from earlier is the most ordinary example of all. A common lens correction produces motion illusions in a large population of wearers [50]. Not a laboratory demonstration. Something people put on their face every morning.

What Illusions Are Actually For

Nothing. That is the point, and it is worth ending on.

Illusions are not a feature the visual system has. They are not a bug it has either. They are what happens at the edge of a strategy, and the strategy is the thing worth understanding.

The strategy is this. You never receive the world, you receive light. Light underdetermines the world. To act in time, something in you commits to the single most likely arrangement that could have produced the light you got, using everything the system knows about how light and surfaces and objects usually behave. That commitment is not offered to you for review. It is delivered as experience.

Almost always, it is right. You reach for the mug and your hand arrives at the mug. The room stays the same size when you walk across it. A white page stays white indoors and outdoors under wildly different illumination. That constancy takes an enormous amount of computation and you have never once noticed it working.

An illusion is a stimulus engineered so that the most likely arrangement is not the actual one. Nothing else about the system changes. The same machinery runs, the same commitment is made, and it lands somewhere the ruler disagrees with.

There is a useful way to feel the asymmetry here. Nobody has ever built a stimulus that makes the visual system perform better than it normally does. There is no anti-illusion, no figure you can look at that improves your judgement of length or brightness beyond baseline. All the engineering runs in one direction, toward the failure, because the failure is the only place the mechanism is exposed. When something works, it is invisible by definition.

Which brings the argument back to Rogers and his uncomfortable question. If the effect is the normal operation of a working system on an unusual input, in what sense is it an illusion at all? You could just as fairly say that the flat printed page is the illusion, and your visual system is the only thing in the room reporting honestly on what such a pattern of light would usually mean.

That reframing does not make the pictures less fun. It makes them more useful. Every illusion is a place where the machinery becomes briefly visible, and there are not many of those.

The confident wrongness is the most human part of it, and it is not confined to seeing. Your sense of what you know behaves the same way, as does your sense of what you remember, which our article on recognition versus recall works through in detail. In each case a shortcut that is usually right produces an answer that feels exactly as certain when it is wrong.

Next time two identical lines refuse to look identical, try to resist the reflex of thinking you have been fooled. Something in you just told you the most probable truth about a pattern of light. It was working perfectly. Somebody simply built the pattern on purpose.