The Gorilla You Would Have Noticed

Watch a short video of two teams passing basketballs and count the passes made by the team in white. It takes concentration; the players move around each other and the ball changes hands quickly.

Halfway through, a person in a gorilla suit walks into the middle of the screen, stops, faces the camera, beats their chest, and walks off. They are visible for about nine seconds.

Roughly half of viewers do not see them. Not "did not pay much attention to" Do not see, and say so with complete confidence, and frequently accuse the experimenter of having swapped the video when it is played again.

Daniel Simons and Christopher Chabris published this in 1999 and it has been replicated many times. The image lands on the retina; the eye works perfectly. What fails is attention, which turns out to be doing vastly more filtering than anybody experiences it doing.

The companion effect is change blindness, and it is stranger. If a change to a scene happens during a blink, a camera cut, or a brief flicker of the image, observers routinely miss alterations they would notice instantly if they occurred as visible motion. Faces change. Large objects appear and vanish. Background colours shift.

Simons ran a version of this in the real world. An experimenter stopped a pedestrian to ask for directions. While they were talking, two workmen carried a door between them, and during the interruption the experimenter was swapped for a different person: different height, different clothes, different voice. Around half the pedestrians carried on giving directions to a stranger without noticing the substitution.

We do not hold a detailed model of the visual scene in our heads. We hold enough to answer the question we are currently asking, and we resample the world when we need something else. The feeling of a complete, stable, detailed visual world is a user interface rather than a description of what is stored.

Which brings us to the second-order effect, which is the one that actually matters.

Ask people in advance whether they would notice the gorilla, or the swapped stranger, and they overwhelmingly say yes. They are not merely wrong about the world; they are wrong about their own perceptual capacity, in a way the errors themselves cannot correct, because the errors are invisible from the inside. There is no experience of not seeing something.

This combination, systematic failure plus intact confidence, is exactly what makes a whole class of everyday reasoning unreliable. "I would have noticed if there had been a motorbike." "I was looking right at it." "I'd know if my child were being bullied." These are sincere statements about which the speaker has no usable evidence.

The driving case is the best documented. The impairment from phone conversation while driving is attentional rather than manual, which is why hands-free is not a fix: simulator studies and real-world data find similar increases in missed signals, delayed braking and failure to detect hazards for hands-free and handheld calls. Interestingly, talking to a passenger is measurably safer, apparently because passengers can see the road and spontaneously stop talking when things get difficult: a conversational partner who is also a second observer.

Before this reads as a catalogue of human defect, note that the filtering is compulsory rather than lazy. The optic nerve carries about a million fibres, and the world contains far more detail than any system could analyse in full. Selection is not an option a better-designed brain would have skipped; it is what perception is. The interesting question is what the filter prioritises, and the answer is: whatever the current task has defined as relevant.

Which is why expertise does not protect you, and this is the finding I would most like people to take away.

In 2013 Trafton Drew and colleagues gave experienced radiologists a stack of chest scans to search for lung nodules. Into the final scan they inserted an image of a gorilla, forty-eight times larger than a typical nodule, in the corner of the lung.

Eighty-three per cent of the radiologists did not report it. Eye-tracking showed that many of them had looked directly at it.

These are people whose visual search is far better than yours for the thing they are looking for. Training sharpens the filter toward the target, which improves detection of what you are hunting and can make an anomaly outside that category less visible rather than more. Expertise is a narrowing as well as a deepening.

The practical response to all of this is not to try harder, which does not work and has never worked. It is procedural: checklists that force a look at categories you are not currently searching for, second readers who arrive with a different task set, and designing systems on the assumption that the person watching will miss the unexpected thing, because half the time, demonstrably, they will.