Change Blindness: Anti-Example Companion To The Invisible Gorilla

Simons and Levin 1998 Door Study showed about half of people miss a stranger swapping in mid-conversation. Three decades of replication confirmed the result. Change blindness is robust, mechanistically grounded, and quietly demolishes most assumptions designers and product teams make about discovery.

Most of the cognitive psychology celebrated in TED talks did not survive the replication crisis. Power posing collapsed. Ego depletion failed preregistration. Bargh’s elderly priming evaporated. Bandura’s Bobo doll generalized further than the data allowed. Mehrabian’s “7-38-55 rule” became a meme detached from its original constrained finding. The track record is so bad that the appropriate prior on any famous psychology demonstration is skeptical-by-default.

Then there is the other category. Findings that look like classroom magic tricks but turn out, when you go looking, to be mechanistically grounded, repeatedly replicated, and consequential for applied work. The Simons and Chabris 1999 invisible gorilla is one. Change blindness — its sister phenomenon, established a year earlier in the same lab — is another.

This article treats change blindness as an anti-example in this hub’s framework: a phenomenon so robust across methodologies and so well-integrated into a coherent theory of perception that it serves as a calibration anchor. If you want to know what a real cognitive-psychology finding looks like, it looks like this. And if you design products, write copy, lay out interfaces, or run anything that depends on users noticing changes in their visual field — you need to internalize the result, because the default assumption “they’ll see the new feature/banner/CTA” is almost certainly wrong.

The Door Study Cold Open

In 1998 Daniel Simons and Daniel Levin published one of the most viscerally surprising demonstrations in modern perception research. The paper is Simons, D. J., and Levin, D. T. (1998). “Failure to detect changes to people during a real-world interaction.” Psychonomic Bulletin and Review, 5(4), 644–649. DOI: 10.3758/BF03208840.

The setup was a piece of guerrilla experimental theater on the Cornell campus. A confederate experimenter, carrying a campus map, approached unsuspecting pedestrians and asked for directions to a specific building. The pedestrian — a real person who had no idea they were in a study — would begin giving directions. Mid-conversation, two other confederates carrying a large wooden door walked rudely between the experimenter and the pedestrian. Behind the cover of the door, the original experimenter switched places with a second experimenter who had been hiding behind the door. The two experimenters were of similar age and gender but were not lookalikes. They wore different clothes, had different builds, different voices, and different hair. The second experimenter then continued the conversation as if nothing had happened.

The question was: did the pedestrian notice that the person they were talking to had been swapped for a completely different person?

Approximately half did not. Out of fifteen pedestrians in the original study, seven failed to notice the substitution and continued giving directions to a stranger they had never seen before. When asked at the end of the interaction whether they had noticed anything unusual about the conversation, those seven said no. Only when explicitly told that the person they had been talking to had been swapped did the change become apparent — and several subjects insisted that the swap could not possibly have happened, because surely they would have noticed.

The hit rate of the failure-to-detect — roughly half of subjects, in a real-world setting, swapping the actual identity of a conversation partner, with everything else held constant — is so high that no plausible methodological critique can dismiss it. This is not a marginal effect that disappears under preregistered replication. It is a binary failure rate, replicated across follow-up studies with different confederate pairs, different locations, different interruption objects, and different cover stories, and the effect appears reliably at roughly the same magnitude.

Simons and Levin called the phenomenon “change blindness,” extending earlier terminology, and the paper became a foundational citation in what became a decade-long research program at the intersection of attention, perception, and consciousness.

What Change Blindness Actually Is

Change blindness is the systematic failure to detect changes in a visual scene when the change occurs during a brief visual disruption. The disruption can be: a saccade (an eye movement), a blink, an occlusion (something briefly blocking the view), a cut between frames in a video, a flicker between two alternating images, or — as in the Door Study — a physical interruption of the line of sight.

The phenomenon is not the same as inattentional blindness. Inattentional blindness — the Simons and Chabris 1999 gorilla finding — is the failure to notice an unexpected event that is continuously present in the visual field but unattended. Change blindness is the failure to notice that something that was present has been replaced or removed. The two phenomena are sister effects: both reveal that the conscious visual experience of the world is sparser and more selective than the rich, complete, continuous experience that introspection suggests.

The mechanism is, roughly, that the visual system does not maintain a stable, persistent representation of the entire visual scene. What feels like a complete, photographic, ongoing perception of the world is in fact constructed moment-to-moment from a much sparser sampled representation. When the visual stream is disrupted — when there is a brief gap during which the scene could in principle have changed — the system does not maintain enough information to detect that change unless attention happened to be focused on the changing element at the time.

This is mechanistically grounded in the basic architecture of the visual cortex, the limits of working memory for visual detail, and the predictive structure of perception. It is not a quirk of a particular paradigm. It is a fundamental feature of how vision works, and it has been demonstrated across dozens of methodologies.

The Flicker Paradigm

The most rigorous laboratory paradigm for studying change blindness was developed by Ron Rensink and colleagues. The foundational paper is Rensink, R. A., O’Regan, J. K., and Clark, J. J. (1997). “To see or not to see: The need for attention to perceive changes in scenes.” Psychological Science, 8(5), 368–373. DOI: 10.1111/j.1467-9280.1997.tb00427.x.

The flicker paradigm presents subjects with two alternating images, separated by a brief blank screen. The original image (call it A) appears for around 240 milliseconds, then a blank gray screen appears for around 80 milliseconds, then a modified image (call it A’) appears for 240 ms, then blank, then A again, and so on, cycling indefinitely. The two images differ in one significant respect — a major object has been added, removed, or changed in color or location. The subject’s task is to detect the change and click on its location as fast as possible.

What Rensink and colleagues found is that subjects often took remarkably long to detect changes that, once pointed out, were visually obvious. Changes to objects of central importance to the scene’s gist were detected within a few cycles. Changes to objects in marginal locations or to background elements often took dozens of cycles — sometimes a full minute of staring at a flickering scene — to detect. And in a substantial fraction of trials, subjects failed to detect the change at all within the test window, even when the changed object occupied a significant portion of the image.

The critical finding was that the blank screen between A and A’ is what produces the blindness. Without the blank screen — when A and A’ alternate directly — the change is detected immediately, because the change itself generates a motion or transient signal that captures attention. The blank screen masks that transient, so the change becomes invisible unless attention happens to be on the right part of the scene.

This was theoretically important because it dissociated the visual change itself from the visual attention required to perceive it. The change was just as physically present in both conditions. What differed was whether the visual system had access to a transient cue. When the cue was absent — when the change happened during a disruption — perception failed.

The flicker paradigm has been replicated in hundreds of studies, used in clinical research on attention disorders, applied to driving simulators and aviation displays, and refined into a standard tool for studying selective attention. It is one of the most reliable paradigms in cognitive psychology.

Mudsplashes — The Severability Test

A standard concern with the flicker paradigm is that the blank screen is artificial. Real-world visual scenes do not contain blank gray screens between frames. So the question arises: does change blindness happen in more naturalistic conditions, or is it an artifact of the flicker manipulation?

The answer came from a clever paradigm published in Nature in 1999. The paper is O’Regan, J. K., Rensink, R. A., and Clark, J. J. (1999). “Change-blindness as a result of ‘mudsplashes’.” Nature, 398(6722), 34. DOI: 10.1038/17953.

In the mudsplash paradigm, subjects view a static scene that periodically has small, briefly-presented dark blobs (“mudsplashes”) appear in random locations on the screen. The mudsplashes are visually distracting but do not occlude the part of the scene where the actual change occurs. During the mudsplash event, a change is made to some other part of the scene — an object is moved, added, or removed.

Subjects in this paradigm showed the same change blindness as in the flicker paradigm, even though there was no blank screen and the changing region was never occluded. The mudsplashes acted as visual transients that captured attention and masked the much smaller transient produced by the actual scene change. The mudsplashes were sufficient to disrupt change detection without ever occluding the change itself.

This was the severability proof. Change blindness was not an artifact of a particular paradigm. It was a generalizable property of how attention and perception interact whenever a transient signal — any signal — disrupts the visual stream. The Nature paper showed that the phenomenon could be produced with arbitrary distractor stimuli that bore no physical relationship to the changing object, which collapsed the space of plausible methodological critiques.

The mudsplash experiment is the kind of methodological work that distinguishes durable science from fragile findings. The authors anticipated the obvious objection (“but blank screens are artificial”), designed a paradigm that addressed it directly, and demonstrated that the effect survived. This is exactly the pattern that, in the rest of this hub, you keep not seeing when you go looking for it.

The 2000 Synthesis Review

By 2000, the literature on change blindness was large enough to warrant a systematic synthesis. The review is Simons, D. J. (2000). “Current approaches to change blindness.” Visual Cognition, 7(1–3), 1–15. DOI: 10.1080/135062800394658.

Simons used the review to organize the rapidly growing literature into a coherent theoretical framework. The major findings he synthesized:

1. The phenomenon generalizes across stimuli. Change blindness has been demonstrated for naturalistic photographs, abstract arrays, motion pictures, real-world environments, and synthetic computer-generated scenes. The effect is not specific to any particular type of visual material.

2. It generalizes across change types. The changes that go undetected include object additions, deletions, color shifts, location shifts, identity substitutions, and orientation changes. The effect is not specific to any particular type of visual modification.

3. It generalizes across disruption types. The intervening disruption can be a saccade, a blink, a film cut, a mudsplash, a flicker, an occlusion, a physical interruption (the Door Study), or a temporal gap in a series of images. The effect does not require a specific kind of mask.

4. Attention is the necessary condition for change detection. Across all of these manipulations, the consistent finding is that change is detected when attention is directed to the changing element at the time of change and missed when attention is elsewhere.

5. The phenomenon is theoretically diagnostic. The fact that observers fail to detect changes to large, central, semantically significant elements of a scene constrains theories of visual representation. Specifically, it rules out the strong form of the claim that perception involves the construction of a complete, persistent, detail-rich internal model of the visual world.

The Simons 2000 review is now itself a foundational citation, accumulating well over a thousand citations across cognitive psychology, vision science, applied human factors, and consciousness research. The framework it laid out has been refined but not overturned in the twenty-six years since.

Applied Implications

The reason change blindness matters beyond cognitive psychology is that the visual world we design, sell, drive in, and witness is full of changes we are not noticing.

Eyewitness Testimony

The most consequential applied domain is eyewitness memory and identification, an area where Elizabeth Loftus’s work over decades has documented systematic failures of recall, recognition, and source monitoring. (See The Loftus eyewitness memory program in this hub.)

Change blindness adds a particularly disturbing layer to that body of work. If half of pedestrians fail to notice when their conversation partner is swapped for a different person, the implications for eyewitness identification are immediate. A witness who saw a perpetrator briefly — across a counter, during an interruption, in a crowd — and then was asked to identify the perpetrator from a lineup might be confidently identifying a person who is not the person they actually saw, because the visual system did not maintain a reliable representation of the original face across the interruption.

This is not a hypothetical concern. The misidentification rate in eyewitness studies is high enough, and the consequences of misidentification severe enough, that the change blindness literature directly informs how police lineups are now structured in many jurisdictions (sequential rather than simultaneous, double-blind administration, confidence calibration at first identification rather than at trial).

Driver Attention

A second applied domain is driver attention, particularly under conditions of distraction. The relevant question is not “can a driver detect a sudden, salient hazard,” because they generally can. The relevant question is “can a driver detect changes to the driving scene during a brief attentional interruption — checking a mirror, looking at a phone, reading a sign — when the change involves an object that becomes a hazard.”

The research is unambiguous: change blindness produces detection failures in driving scenarios at rates that translate directly into accident probability. A pedestrian who steps into the road during the moment a driver glances at their phone is, perceptually, in roughly the same situation as a confederate swapped in behind the door. The driver who returns their gaze to the road has no transient signal alerting them to the change. The pedestrian is not seen.

This is the perceptual substrate of the driving-while-phoning effect. It is not that the driver is “not paying attention” in a generic sense — most drivers are paying attention to the road, intermittently. It is that the brief disruptions during which they are not looking at the road create exactly the conditions under which change blindness produces detection failure.

UX and Interface Design

The third applied domain, and the one most relevant to anyone working in product or marketing, is interface design. The default assumption embedded in most product launches and most landing-page redesigns is that users will notice the new feature, the new banner, the new CTA, the redesigned navigation. They will, the assumption goes, encounter the change in the course of their normal usage, and the change will be salient enough to register.

This assumption is almost always wrong. The conditions under which a user encounters a redesigned interface are exactly the conditions under which change blindness operates: brief attentional disruptions (a notification, a page reload, a context switch), no transient cue connecting the old version to the new version, and a primary task (find the thing they came for) that monopolizes attention.

The empirical literature on “banner blindness” — the systematic failure of users to notice banner ads, sidebar promotions, and similar peripheral content — is a special case of change blindness combined with learned suppression. (See Banner blindness in this hub for the dedicated treatment.) The core phenomenon, though, is the change blindness substrate: users are not maintaining a detailed representation of the interface, so changes to peripheral or non-task-relevant elements do not register.

The implication for product teams is that “we launched a new feature and users will discover it” is a hypothesis, not a fact. Discovery must be tested, instrumented, and verified. The discovery rate of any new feature should be measured against an empirical baseline, not assumed to match the team’s enthusiasm for it.

Inattentional Blindness — The Sister Phenomenon

The connection between change blindness and the more famous inattentional blindness paradigm is close enough that the two literatures effectively form a single research program. Both Simons-Levin 1998 (change blindness, Door Study) and Simons-Chabris 1999 (inattentional blindness, gorilla video) were produced in the same lab in adjacent years, and the theoretical interpretation of both findings runs along the same lines.

The distinction matters, though. Inattentional blindness is the failure to detect an unexpected event in the visual field while attention is focused on a primary task. The gorilla is continuously visible for nine seconds; the failure is not that it disappears, but that it never enters consciousness in the first place. Change blindness is the failure to detect a modification to the visual scene across a brief disruption. The change does enter the scene; the failure is detecting the difference between before and after.

Together the two phenomena bound the gap between the rich, complete, continuous visual experience we introspectively believe we have and the sparse, selective, attention-mediated reality of perception. (See Inattentional blindness — The Invisible Gorilla for the companion article.)

Predictive Coding And The Modern Synthesis

The modern theoretical framework that absorbs both change blindness and inattentional blindness is predictive coding — the view that perception is fundamentally a process of generating top-down predictions about the visual world and reconciling those predictions against bottom-up sensory data. The canonical formulation is Clark, A. (2013). “Whatever next? Predictive brains, situated agents, and the future of cognitive science.” Behavioral and Brain Sciences, 36(3), 181–204. DOI: 10.1017/S0140525X12000477.

The predictive-coding interpretation of change blindness goes roughly as follows. The visual system does not store a detailed picture of the world from moment to moment. Instead, it maintains a generative model — a set of predictions about what should be in the scene — that is updated by sensory evidence. Detection occurs when sensory evidence contradicts predictions; non-detection occurs when sensory evidence is consistent with predictions or when no evidence is available.

A change during a brief disruption produces no evidence, because the disruption masks the transient that would normally signal change. The post-change scene is consistent with the predicted scene (a stranger talking to me about directions; a basketball game with the players continuing to play; a road with cars and pedestrians on it), so the predictions are not updated. The change does not register because the system has no signal to update on.

This framework has explanatory power well beyond change blindness. It accounts for the gestalt of conscious visual experience (it feels complete because the model is complete, not because the data are complete); the limits of working memory for visual detail; the role of expectation in perception; and a wide range of clinical phenomena involving disrupted perception. It is one of the more successful theoretical syntheses in cognitive science of the last twenty years.

The predictive-coding interpretation matters here for two reasons. First, it gives change blindness a coherent mechanistic story, which is part of what makes the phenomenon resistant to the methodological critiques that destroyed so many other psychology findings — there is a worked-out theory that explains why the effect would occur, not just a paradigm where it does. Second, it bridges change blindness into a research program that is still active and productive in 2026, which means the construct is not a freeze-dried 1990s artifact but an ongoing topic of inquiry.

The Strategist Takeaway — Test Discovery, Don’t Assume It

If you are running a product, designing an interface, writing marketing copy, or building anything that depends on users noticing things, the change blindness literature collapses to one operational rule: discovery is empirical, not assumed.

The default mental model most teams operate from is some version of the rich-percept assumption — users have a detailed mental picture of your product, they will notice when you change things, they will see the new feature when it appears, they will read the banner you placed at the top of the page, they will discover the secondary CTA you added below the fold. The assumption is so embedded in product workflows that it is rarely articulated, let alone tested.

The empirical reality is that none of those things are likely to happen at the rate the team is assuming. Discovery rates for new features in mature products are routinely below 10%. Click-through rates on prominently placed CTAs are routinely below 5%. Change-detection rates for redesigned interface elements are routinely below 50% even among heavy users. These are not failures of the users. They are the predictable output of change blindness operating in exactly the conditions you have designed for.

The operational response is three-part:

First, instrument discovery. Every new feature should have a discovery-rate metric, measured directly, against a defined baseline. Do not infer discovery from downstream usage; many features get used by the small fraction of users who discover them, and the headline usage number conceals the discovery problem entirely.

Second, treat in-product education as load-bearing, not optional. If you cannot get a meaningful fraction of your users to notice a feature through ambient placement alone, the feature will not earn its development cost. Tooltips, onboarding, contextual nudges, and product tours are not “nice to have” UX polish; they are the mechanism by which you compensate for change blindness in the underlying perceptual system.

Third, A/B test the discovery, not just the conversion. Most experimentation programs measure the conversion rate from feature-encountered to feature-used. They do not measure the prior rate of feature-encountered itself, which is the dominant variance in most launch outcomes. A feature with 90% discovery and 10% conversion is a stronger launch than a feature with 10% discovery and 90% conversion, and they look identical on a conversion-rate dashboard.

The deeper point — the strategist point that the entire replication-crisis hub keeps converging on — is that human perception, judgment, and behavior are systematically less reliable than the folk theory of them, and the only correction is measurement. The same epistemic move that protects you against believing in power posing or ego depletion protects you against assuming your users will see your new banner. The discipline is the same one. Run the test. Believe the result. Update.

Sources

• Simons, D. J., and Levin, D. T. (1998). “Failure to detect changes to people during a real-world interaction.” Psychonomic Bulletin and Review, 5(4), 644–649. DOI: 10.3758/BF03208840.
• Rensink, R. A., O’Regan, J. K., and Clark, J. J. (1997). “To see or not to see: The need for attention to perceive changes in scenes.” Psychological Science, 8(5), 368–373. DOI: 10.1111/j.1467-9280.1997.tb00427.x.
• Simons, D. J. (2000). “Current approaches to change blindness.” Visual Cognition, 7(1–3), 1–15. DOI: 10.1080/135062800394658.
• O’Regan, J. K., Rensink, R. A., and Clark, J. J. (1999). “Change-blindness as a result of ‘mudsplashes’.” Nature, 398(6722), 34. DOI: 10.1038/17953.
• Clark, A. (2013). “Whatever next? Predictive brains, situated agents, and the future of cognitive science.” Behavioral and Brain Sciences, 36(3), 181–204. DOI: 10.1017/S0140525X12000477.
• Simons, D. J., and Chabris, C. F. (1999). “Gorillas in our midst: Sustained inattentional blindness for dynamic events.” Perception, 28(9), 1059–1074. DOI: 10.1068/p281059.

Related In This Hub

• Inattentional Blindness / The Invisible Gorilla — The sister paradigm. Same lab, adjacent year, same theoretical synthesis. The two phenomena bound the gap between the rich-percept folk model and the sparse-percept reality.
• Loftus Eyewitness Memory — The applied domain where change blindness most directly bears on consequential decisions. If conscious perception is this unreliable across brief interruptions, eyewitness identification is operating on a much thinner evidentiary base than the legal system has historically assumed.
• Banner Blindness — The interface-design special case. Banner blindness is partly learned suppression and partly change blindness; the underlying perceptual substrate is the same.
• Mehrabian 7-38-55 Rule — Companion failure case. Mehrabian is what happens when a constrained finding gets popularized as a general perceptual claim. Change blindness is what happens when a robust finding gets popularized accurately.
• Mere Exposure Effect — A related case of a perceptual finding that survived replication scrutiny. Repeated exposure produces preference shifts; the construct is real even when the magnitudes have been overclaimed in popular accounts.

FAQ

Did Simons and Levin’s Door Study replicate?

Yes, and at multiple scales. The original 1998 paper itself reported the result across more than one variant of the paradigm. Subsequent work by Simons and other groups extended the result with different confederate pairs, different cover stories, different physical interruptions, and different real-world settings. The basic effect — roughly half of subjects failing to notice a substantial identity change to a conversation partner during a brief interruption — has been repeated reliably for more than two decades. This is not a one-off demonstration; it is a robust experimental finding.

Is change blindness the same thing as inattentional blindness?

No, but the two are closely related and often confused. Inattentional blindness is the failure to notice an unexpected event that is continuously present in the visual field while attention is engaged on a primary task — the canonical example is the gorilla in the Simons-Chabris 1999 video. Change blindness is the failure to detect that something has changed between two views of the same scene, across some intervening disruption — the canonical example is the Door Study. Both phenomena reveal limits of conscious visual perception, but the mechanisms are distinct. Inattentional blindness is about attention allocation during continuous viewing; change blindness is about visual representation across disruption.

Why didn’t change blindness fail like so much else in cognitive psychology?

Several reasons converge. The effect size is large — half of subjects, not 5% of subjects. The phenomenon is self-replicating across paradigms — flicker, mudsplash, real-world interruption, film cut, occlusion, saccade — which would not be the case if the original finding were a methodological artifact. The theoretical framework is mature and mechanistically grounded in the architecture of visual attention. And the demonstration is accessible to direct replication: any lab, any classroom, any individual can run a change-blindness paradigm and observe the effect. None of these conditions held for the social-priming and ego-depletion findings that collapsed under scrutiny.

Does change blindness apply to people who know about change blindness?

Yes. Knowing that change blindness exists does not protect you from experiencing it. The phenomenon operates at a level of visual processing that is largely insulated from explicit knowledge or top-down attention. You can be a vision scientist who has studied change blindness for thirty years and still fail the Door Study yourself. This is one of the more humbling aspects of the literature, and it is part of why the operational rule for product and design work is to measure discovery directly rather than relying on team intuition about what users will notice.

What is the most underrated applied implication?

Probably the consequences for institutional decision-making in domains where multiple people view the same scene and are expected to agree on what they saw. Eyewitness identification is one. Medical imaging review — radiologists looking at scans for tumors — is another; the Drew, Vo, and Wolfe 2013 study extended inattentional-blindness paradigms to expert radiologists and found very high miss rates for unexpected anomalies, with direct implications for diagnostic protocols. Aviation cockpit monitoring is a third. Security and surveillance design is a fourth. In any institutional setting where the operating assumption is “if it happened in plain sight, someone will have noticed,” change blindness and inattentional blindness together suggest the assumption is wrong at a rate that is operationally significant.

Does change blindness undermine the reliability of human perception generally?

Not exactly, and the popular framing here often overstates the case. Human perception is highly reliable for the things it is optimized for: detecting motion, recognizing faces under normal conditions, navigating complex environments, identifying objects of immediate behavioral relevance. What change blindness shows is that perception is selectively reliable — high-fidelity for attended, task-relevant, predicted aspects of the visual world and low-fidelity for unattended, peripheral, or unexpected aspects. The conscious experience of seeing a complete, detailed, persistent world is partly construction, not entirely recording. This is a calibration about the limits of perception, not a global indictment of it.

Where should I go to actually experience change blindness?

The original Door Study video is available online and is striking to watch. The Rensink flicker paradigm has been adapted into many freely available web demos — pairs of alternating images with one change that you have to find, often taking many seconds even when the change is large. Both are worth seeing firsthand, because the perceptual experience of failing to detect an obvious change is more convincing than any written description of the phenomenon. Once you have experienced it yourself a few times, the implications for any system that depends on human visual monitoring become much more concrete.