The Invisible Gorilla – Inattentional Blindness and the Limits of Attention

The Hook

Picture yourself in a code review. In front of you is a pull request with forty changed files. Your brief is clear: check whether the new retry logic is correct. You work carefully, line by line, and you do find two bugs in the retry logic. Good work.

Three weeks later it turns out that the same pull request wrote an API key in plain text into a configuration file. You had the file open. The diff was highlighted in green, right in your field of view. You looked at it – and saw nothing.

The obvious explanation is inattention, sloppiness, too little time. The psychologically more interesting explanation is this: it is not a failure of your diligence but a property of your visual system. And it can be produced in the lab so reliably that it reduces lecture halls to stunned laughter.

The classic demonstration of this effect is a gorilla.

The Experiment That Became a Legend

In 1999 Daniel J. Simons and Christopher F. Chabris, both then at the Department of Psychology at Harvard University, published a paper in the journal Perception titled "Gorillas in our midst: sustained inattentional blindness for dynamic events" (volume 28, issue 9, pages 1059–1074).

The setup is quickly told. Observers watched a 75-second video. In it, six people played basketball, three in white shirts and three in black. The task: count the passes made by one of the two teams. After 44 to 48 seconds an unexpected event entered the scene – either a tall woman with an open umbrella or a shorter woman in a gorilla costume that fully covered her body. The event lasted five seconds; the players carried on with their actions during and after it.

Five seconds is not a subliminal flash. Five seconds is an eternity.

Afterward, observers were questioned in a cascade of increasingly direct questions, starting with whether they had noticed anything unusual about the video, and ending with the specific question of whether they had seen a gorilla, or a woman carrying an umbrella, walk across the screen.

The result that has appeared in every introductory course ever since: 46 percent of observers did not notice the unexpected event. 54 percent did.

This is where most retellings stop. That is a shame, because the genuinely interesting insight lies in the structure underneath.

What the Numbers Actually Say

Simons and Chabris did not run a single experiment but one experiment with sixteen conditions. Their design was four-factorial (2×2×2×2) with twelve people per cell. Of 228 recruited observers, 192 remained in the analysis after exclusions – those excluded included fourteen people who already knew the phenomenon and nine who lost count.

The four factors were:

This table holds the real lesson. The effect is not a fixed percentage but a function. It grows stronger when the task demands two things at once, and it grows dramatically stronger when the unexpected object is dissimilar to the attended objects.

The last point is the subtlest. Those counting passes by the team dressed in black noticed the black gorilla in 58 percent of cases. Those counting passes by the team dressed in white noticed it in only 27 percent of cases. Same scene, same object, same location on the screen – the difference lies solely in which property attention has currently set as relevant. From this the authors conclude that spatial proximity of the unattended object to attended locations does not appear to affect detection, and that observers therefore attend to objects and events, not to spatial positions.

In one of the extreme cells of the design – transparent display, gorilla, easy counting task, white team attended – the detection rate was eight percent. Twelve people, one of whom saw it.

In fairness to perception, the other side deserves mention as well: even in the more realistic opaque display, in which all seven actors were filmed simultaneously and could occlude one another, 33 percent still missed the event. And the authors themselves point out that the umbrella woman and the gorilla differ in more than one property – color among them – so attributing the difference solely to "similarity to the attended team" should be treated with caution. They likewise note that the transparent superimposed displays do not resemble real-world scenes.

The Door: The Same Effect Without a Screen

A year earlier, Daniel J. Simons and Daniel T. Levin had shown that no screen is needed. Their paper "Failure to detect changes to people during a real-world interaction" appeared in 1998 in Psychonomic Bulletin & Review (volume 5, issue 4, pages 644–649).

On the campus of Cornell University an experimenter holding a campus map approached pedestrians and asked for directions. During the conversation, two further experimenters carried a door between the pair. For roughly one second the view was blocked. In that second the first experimenter grabbed the back of the door and walked off; the person who had been carrying the door stayed behind and continued asking for directions – holding an identical copy of the map.

The two experimenters wore different clothing, differed by about five centimeters in height, and had clearly distinguishable voices.

Of fifteen pedestrians approached, seven reported noticing the change. Eight out of fifteen did not notice that the person they were speaking to had been swapped.

Caution is called for here, and the authors model it well. Fifteen people is a very small sample; the paper reports no statistical tests. Notably, everyone who noticed the switch was roughly the age of the experimenters (about 20 to 30 years old), while the non-noticers were roughly between 35 and 65. The authors explicitly offer age as one possible explanation, not as a finding – they give no exact counts per age group.

A second experiment tested the hypothesis derived from this, that membership in one's own social group is the deciding factor. This time both experimenters appeared dressed as construction workers, and only younger pedestrians were approached. Of twelve people, only four noticed the switch – the same age group that had succeeded in every case in the first experiment.

The authors also discuss the obvious counter-hypothesis that pedestrians did notice the change but politely did not mention it. They consider this unlikely: expectancy effects in such an interview situation would tend to produce more reports of a switch, not fewer. Conversely, in the second experiment they discounted three people who claimed a switch only when asked directly, yet could not describe any difference between the two experimenters. A precise analysis of eye contact was not possible, because the camera's vantage point did not allow it.

When It Gets Serious: The Kenny Conley Case

Up to this point this is an amusing laboratory curiosity. In 2011 it became a question of some weight.

On the night of 25 January 1995, at around two in the morning, Boston police officer Kenny Conley was chasing a shooting suspect. Shortly before, the undercover officer Michael Cox had arrived; other officers mistook him for a suspect, assaulted him from behind, and beat him. Conley chased his suspect over a chain-link fence and later testified that he had run right past the spot where Cox was being attacked but had not seen the incident.

Christopher F. Chabris, Adam Weinberger, Matthew Fontaine, and Daniel J. Simons describe, in "You do not talk about Fight Club if you do not notice Fight Club: Inattentional blindness for a simulated real-world assault" (i-Perception, volume 2, issue 2, pages 150–153, 2011), the logic that investigators, prosecutors, and jurors followed at the time: because Conley could have seen the beating, he must have seen it – and was therefore lying to protect his colleagues. Conley was found guilty of perjury and obstruction of justice and received a sentence of thirty-four months. The authors record that apparently none of the principals in the case considered that Conley might have been telling the truth.

The authors had 94 college students run a 400-meter loop on the Union College campus across three studies. The task: follow a confederate running ahead at a distance of about nine meters and count how often he touched his head. About 125 meters into the route, three further confederates staged a fight eight meters off the path, in which two beat up a third – with shouting, grunting, and coughing. At night the scene was visible for at least fifteen seconds before the participant passed it; in daylight it became visible about twenty seconds into the run and stayed visible for at least thirty seconds.

The difference between high and low cognitive load within study 3 was statistically significant (χ²(1) = 5.03, p = 0.02). So even without an additional task, a good quarter of participants missed a staged fight eight meters away.

What matters is how restrained the authors are in phrasing their result. They write that no scientific study can prove or disprove a particular cause of a specific historical event; their results show only that Conley's claim that he did not see the beating of Michael Cox might well have been truthful. They further point out that their participants – unlike Conley – knew they were taking part in research and had been assigned a task requiring focused attention, and that physical exertion during the run was not controlled. More generally, the authors note, generalization from studies using videos or computerized displays to real-world events is not necessarily valid.

One footnote matters for the legal history: Conley later won an appeal, and the government decided not to retry him. The authors state explicitly that the inattentional blindness explanation did not contribute to either of these decisions.

The Gorilla in the Chest CT

The most uncomfortable study in this series comes from radiology. Trafton Drew, Melissa L.-H. Võ, and Jeremy M. Wolfe of the Visual Attention Lab at Harvard Medical School and Brigham and Women's Hospital published "The Invisible Gorilla Strikes Again: Sustained Inattentional Blindness in Expert Observers" in Psychological Science in 2013 (volume 24, issue 9, pages 1848–1853).

Twenty-four radiologists performed a task familiar to them: searching for lung nodules in chest CT scans. Into the last case presented, the researchers had inserted the image of a gorilla – forty-eight times the size of an average nodule.

Eighty-three percent of the radiologists did not see the gorilla. Eye tracking revealed that the majority of those who missed it had looked directly at its location.

The authors' conclusion in the abstract is terse: even expert searchers, operating in their own domain of expertise, are vulnerable to inattentional blindness.

Two qualifications belong with this. First: the full text of this paper was not accessible for the present article; the figures given here come from the publicly available abstract on the publisher's page, which also means the limitations stated by the authors themselves are missing. Second – and more important – the study does not measure the radiologists' diagnostic competence. It measures whether an object irrelevant to the task is noticed. A radiologist who looks for nodules and finds nodules is doing the job; the gorilla is precisely what the configured attentional filter was not built for. The study shows that expertise does not confer immunity to inattentional blindness – not that expertise is worthless.

Does Knowing the Trick Help?

An obvious hope: anyone who knows about the gorilla is forearmed. Daniel J. Simons tested this in 2010 in i-Perception (volume 1, issue 1, pages 3–6) under the title "Monkeying around with the Gorillas in Our Midst: Familiarity with an Inattentional-Blindness Task Does Not Improve the Detection of Unexpected Events".

He showed people who already knew the original video a new video in which, besides the gorilla, two further unexpected events occurred: a curtain changed color, and one player left the scene. The result according to the abstract: those forewarned spotted the gorilla consistently – but, in the abstract's wording, were "slightly less likely" to notice the other events.

Simons's conclusion in the abstract is, in substance, that foreknowledge that unexpected events might occur does not enhance the ability to detect other such events.

Here too a qualification: the full text was not accessible, so participant counts and significance statements could not be checked. The phrase "slightly less likely" in the abstract leaves open whether the difference was statistically significant and against which comparison group it was determined. The defensible claim is therefore the negative one – foreknowledge did not help – not the stronger claim that foreknowledge hurts.

What We Do Not Know

It would be dishonest to end this article without showing the open flanks. The following points are contested in the specialist literature; the overview I rely on here is the English-language Wikipedia article "Inattentional blindness", which names the respective primary works. It is explicitly not a primary source.

Seeing or remembering? The popular reading is that the object is never consciously perceived. A competing explanation is called inattentional amnesia: the object is indeed seen but forgotten so quickly that it is no longer available when the question is asked. This debate is methodologically hard to settle, because an advocate of the memory account can always object that the memory test came too late. The overview concludes that both mechanisms are likely involved.

Who is susceptible? The conjecture that people with high working memory capacity are less susceptible has produced contradictory findings: work by Seegmiller, Watson, and Strayer (2011) reports a strong relationship, while work by Bredemeier and Simons (2012) and by Kreitz, Furley, and Memmert (2015) finds none. The state of the evidence is considered inconclusive.

How far does the lab carry? That an effect occurs in a video and on a campus running route does not automatically mean it explains the performance of professionals under real pressure to the same degree. Chabris and colleagues state this reservation themselves. And the overview notes the objection raised against using inattentional blindness as a defense argument: that it could be used to defend nearly any police shooting.

I am of the opinion that this last point marks the real practical challenge: the effect is real enough to design processes around, and at the same time unspecific enough to serve as an excuse. Anyone who takes it seriously should therefore build it into systems – not into apologies.

A Small Framework for Practice

A rough risk grid can be derived from the verified findings. The middle column states the finding; the right-hand column states my derived recommendation, which was not tested in any of the studies discussed. The separation matters, so that the whole table is not read as evidence-based.

The Central Takeaway

The usable core is not "people are inattentive". The usable core is: attention is a filter that you set unconsciously when you formulate your task – and whatever does not pass the filter is not available to you afterward, even if your eyes were pointed directly at it. Whether it never reached awareness or was immediately lost again is an open question in research; for practical purposes it makes no difference.

From this follows a concrete course of action, and one that is not "try harder". Simons's 2010 findings argue, at minimum, against the idea that mere foreknowledge of possible unexpected events improves the detection of other such events.

What follows instead: separate your checking tasks rather than bundling them. The next time you face a code review, a log analysis, or a contract review, define one search brief per pass and go over it a second time with a deliberately different search brief – not "look closely again", but with a different filter. And build an explicit entry into your checklist for the category "things nobody was looking for".

It should be noted that this very procedure was not tested in any of the studies discussed here; it follows from the mechanism, not from a finding. What is established is only that the configured search brief massively changes the detection rate. I am of the opinion that the consequence is to make the search brief itself an object of process design – but that is a derivation, not a measurement.

Reflection Question

In which of your routine checking processes – code review, monitoring, acceptance testing, budget control – is the search brief so narrow and so constant that for years nobody has had a chance to find anything that is not on the list?

Fact Check

Automated AI check against primary sources; no human expert review.

Roughly 45 checkable claims (hard facts, quotations, source attributions, and synthesis claims) were checked against the sources opened during this run; 14 were corrected or qualified. Of the five primary sources, two (Drew et al. 2013, Simons 2010) were accessible only as abstracts; this is noted in the text at the relevant places.

The three most important corrections:

1. The figure for the eight-percent cell in Simons & Chabris (1999) was misassigned in the draft: it belongs to the easy counting condition (transparent, gorilla, white team attended), not the hard one. In addition, "the most extreme cell" was weakened to "one of the extreme cells", since the values of the remaining cells are not fully documented here.
2. The recommendation "a second pass with a different search criterion" originally stood in the evidence column of the risk table and thus read as a study result. It is a derivation from the mechanism and was moved into a separate column explicitly marked as untested.
3. "The Central Takeaway" claimed that what is missed "does not exist" – a purely perceptual reading that contradicted the section on the debate between perceptual and memory failure. It was reduced to the observable claim that the object is not available afterward.

Further corrections concerned: the fifteen-second visibility figure, which applies only to the night condition; the rendering of the authors' argument against the politeness hypothesis in Simons & Levin (1998); the generalization of the criticism from "police shooting" to "police misconduct"; the claim of a comparison with "unprepared" participants in Simons (2010); the description of the sixteen conditions as "sixteen experiments"; and the inclusion of "exhaustion" as a documented risk factor, even though physical exertion was explicitly not controlled in the study.

Not verifiable from primary sources were the statements on the working-memory debate and on the criticism of the forensic use of the effect; they are explicitly attributed in the text to the overview source named.

Sources and Further Reading

1. Simons, D. J., & Chabris, C. F. (1999). Gorillas in our midst: Sustained inattentional blindness for dynamic events. Perception, 28(9), 1059–1074. https://doi.org/10.1068/p281059 — full text: http://www.chabris.com/Simons1999.pdf
2. Simons, D. J., & Levin, D. T. (1998). Failure to detect changes to people during a real-world interaction. Psychonomic Bulletin & Review, 5(4), 644–649. https://doi.org/10.3758/BF03208840
3. Chabris, C. F., Weinberger, A., Fontaine, M., & Simons, D. J. (2011). You do not talk about Fight Club if you do not notice Fight Club: Inattentional blindness for a simulated real-world assault. i-Perception, 2(2), 150–153. https://doi.org/10.1068/i0436
4. Drew, T., Võ, M. L.-H., & Wolfe, J. M. (2013). The Invisible Gorilla Strikes Again: Sustained Inattentional Blindness in Expert Observers. Psychological Science, 24(9), 1848–1853. https://doi.org/10.1177/0956797613479386 — accessible only as an abstract for this article.
5. Simons, D. J. (2010). Monkeying around with the Gorillas in Our Midst: Familiarity with an Inattentional-Blindness Task Does Not Improve the Detection of Unexpected Events. i-Perception, 1(1), 3–6. https://doi.org/10.1068/i0386 — accessible only as an abstract for this article.
6. Wikipedia: Inattentional blindness. https://en.wikipedia.org/wiki/Inattentional_blindness — used as an overview of the state of debate, not as a primary source.

Cross-References in the Vault

• The Predictive Brain: Predictive Processing and the Illusion of Perception — the other side of the same coin: why the brain constructs perception rather than recording it.
• The Invented Memory: Why Recollection Is Not a Recording – and How Elizabeth Loftus Showed You Can Remember Things That Never Happened — relevant to the debate "not seen, or merely not remembered".
• The 38 Witnesses Who Never Were: The Bystander Effect Between Laboratory Finding and Video Evidence — a second case in which a famous story outran the evidence.
• The Peak of Not Knowing: Why Psychology's Most Famous Effect May Be Just a Statistical Mirage — how a robust finding turns into an overstated popularization.
• Desirable Difficulties: Why Effortless Learning Deceives and Memory Lives on Effort — why "more effort" and "better method" are not the same thing.

← All articles