The Waterfall Illusion: Why Still Objects Drift After You Look Away

After two centuries of study the mechanism is still argued over. What is settled is which answers are wrong.

What you are actually seeing

The motion aftereffect — the MAE, in the literature — is a sensation of movement in something that is not moving. After prolonged exposure to motion in one direction, a stationary scene appears to drift in the opposite direction. Downward water produces upward rocks. A spinning spiral produces a scene that seems to shrink or swell.

The crucial detail, and the one almost every popular account skips, is that nothing changes position. The rocks appear to stream upwards and yet stay precisely where they are. You can watch a thing move and remain in place at the same time, which is not a sensation ordinary life has prepared you for.

That is a structural clue about how vision is built, and we will come back to it.

Addams at the Falls of Foyers

The illusion is named for an observation made in 1834 by Robert Addams at the Falls of Foyers, on the shore of Loch Ness. He looked away from the water to the wet rock face beside it and found the rock apparently climbing. He wrote it up, and the waterfall got its name attached to the effect permanently.

He was not first. Aristotle described something recognisably similar around 330 BC, Purkinje rediscovered it in the nineteenth century, and Wohlgemuth produced the first systematic monograph in 1911. Addams gets the credit largely because he gave it a memorable setting.

You do not need a waterfall. Credits scrolling up a screen or a passing train seen from a platform will do.

The explanation everyone gives

Here is the version in almost every textbook summary and science video: your brain has cells tuned to specific directions of motion. Stare at downward motion long enough and the downward cells get tired. With the downward signal weakened, the upward cells win by default, and you see upward motion.

The first half of that is correct. Direction-selective neurons exist, they adapt to sustained motion, and the aftereffect depends on that adaptation. The second half — fatigue — is the part the standard review by Anstis, Verstraten and Mather calls almost certainly incorrect, and it has been known to be shaky for decades.

Several things do not fit. Some direction-selective neurons barely adapt at all, which a fatigue account cannot accommodate. The timing is wrong: recovery does not follow the course that depletion would predict. And when Barlow and Hill recorded from direction-selective retinal ganglion cells in rabbit in 1963, they found that after the motion stopped, firing fell below the cell's own baseline and then climbed back over roughly thirty seconds, with no overshoot. A tired cell goes quiet. This cell went quieter than quiet, which is a different thing, and it is the thing that produces the illusion.

The ratio, not the volume

Sutherland proposed the fix in 1961: perceived direction does not depend on how loudly the downward detectors are firing, but on the ratio between detectors tuned to opposite directions. At rest both sides idle at some baseline, the ratio is balanced, and you see no motion.

Adapt one side and its resting output drops below baseline. The ratio is now genuinely tilted, so the system reports motion — correctly, by its own logic. It is not a tired brain producing a glitch. It is an intact comparison circuit given a dishonest input.

That distinction matters more than it sounds. A fatigue account makes the illusion a malfunction. The ratio account makes it the system working exactly as designed on an input it was never built to receive, which is the same shape of story as the moon looking larger on the horizon — a mechanism doing its job on evidence that happens to be misleading.

The experiment that rules out tired eyes

If adaptation were fatigue in the eye, it would be confined to the eye that did the staring. It is not.

Adapt one eye with the other closed, then test with the adapted eye closed instead, and the aftereffect still appears in the eye that saw nothing — at roughly sixty per cent of full strength. If the test pattern is dynamic rather than static, transfer is essentially complete.

That is decisive. Signals from the two eyes do not combine until the visual cortex, so an aftereffect that crosses between them must involve neurons receiving input from both. And the confirming case is as clean as this literature gets: people who had a squint in childhood, and therefore never developed normal binocular cortical cells, show no interocular transfer at all.

The static-versus-dynamic difference points the same way. A static test picks up adaptation at one stage; a flickering test picks up adaptation at a later, binocular stage. Culham and Cavanagh found that attention-driven motion — perceived motion in an ambiguous flickering grating, with no consistent physical motion present — produced an aftereffect only on dynamic tests. There is more than one site involved, and they are not interchangeable.

Motion that goes nowhere

Now back to the strange part. In the MAE you see movement with no displacement, and that combination should be impossible if motion were computed by tracking positions over time.

It is not. Vision has dedicated motion machinery that operates largely independently of the machinery tracking where things are, and the aftereffect is the clearest everyday demonstration that the two can be pulled apart. Your motion system reports upward; your position system reports stationary; both reports reach you, and neither defers.

This creates a real measurement problem. You cannot cancel a motion signal by drifting the test pattern the other way, because real motion also moves the thing. The workaround is a field of mixed drifting and twinkling dots, with just enough coherent motion added to null the aftereffect — and that technique detects lingering adaptation after the visible illusion has faded.

How long it lasts, and what changes it

A straightforward MAE from a minute of staring usually lasts under a minute itself. Duration grows roughly as the square root of adapting time, so doubling your staring does not double the aftereffect — there are diminishing returns almost immediately.

Attention changes it. Give someone a demanding letter-identification task during the test period and the aftereffect gets shorter, which is hard to square with any purely passive account of adaptation.

And at the far end, contingent aftereffects — where motion is paired with a colour or an orientation — can persist for days or weeks. The same word, "adaptation", is covering processes operating on timescales that differ by four orders of magnitude, which is worth remembering whenever you read a confident single-mechanism explanation.

What is still unsettled

Opponency itself is not fully established. The ratio account predicts that adaptation should shift the point at which a pattern looks motionless, and Raymond and Braddick found no such shift, which is more consistent with independent direction channels than opposed ones. Mather and Harris have proposed an opponent comparator stage downstream of the detectors, which would preserve the ratio logic while explaining why some manipulations produce no aftereffect at all.

So the honest summary is this: adaptation of direction-selective cortical neurons, at several sites, altering a comparison between motion signals. Not fatigue. The precise form of the comparison is still argued about.

It is worth noticing what kind of illusion this is. Nothing is being inferred, assumed or guessed. A sensory channel is reporting motion that is not happening, and the report is accurate to the state of the channel — which is exactly what happens when menthol makes your skin feel cold while the temperature does not move at all.

Producing it yourself

The reliable version takes about ninety seconds and needs no waterfall.

• Find sustained one-directional motion that fills a decent part of your view — scrolling credits, a looping video of falling water, a spinning spiral.
• Fix your gaze on one point and hold it. Letting your eyes track the motion defeats the whole thing, because tracking cancels the retinal motion you are trying to adapt to.
• Keep going for sixty to ninety seconds. Longer helps, but less than you would expect.
• Switch to a stationary textured surface — a wall, a page, a rug. Plain white works poorly; the effect needs something to attach to.
• Notice that the surface drifts the other way and does not actually go anywhere.

If nothing happens, the usual culprit is eye movement during adaptation.