AbilityBenchPerceptual bias
Optical illusion test that hands back a number about you
Three classic figures, nine adjustments, and no gallery: you set one part of each figure until it looks equal to the other, and the page reports the percentage by which your setting misses physical equality. Free, no account, and nothing is held back. Every figure is set twice — once starting far too small, once far too large — and then set once more after the true match has been shown to you, which measures the thing worth measuring: how much of the effect knowledge removes.
- 100% free
- No signup
- 3 figures, 9 settings
- Both starting directions
- Untimed
Three figures, nine adjustments, no pictures to admire. You drag one part of each figure until it looks equal to the other part, and the page reports how far from physically equal you stopped. Each figure is adjusted twice — once starting far too small and once far too large — and then once more after you have been shown the true answer.
What the page does to stop you cheating yourself
- The slider never shows a number, and the travel-to-size mapping is redrawn for every adjustment: a different offset, and a coin flip on which end makes the figure bigger. A control whose middle is the right answer measures your memory of the middle.
- The two halves of each figure get independent random horizontal nudges, so lining an edge up against the panel border is not available.
- The two starting directions are the whole reason there are six blind adjustments rather than three. Settings made from below land short of settings made from above, in every psychophysical task ever run, and the average of the pair is the estimate.
Nothing here is timed. Take as long as you want on each one and go back and forth as often as you like — the setting you confirm is the only thing recorded.
How to measure an illusion instead of looking at one
Set the control, confirm, repeat from the other end, then do it knowing the answer.
Fix your distance and, if you can, measure the screen
How strongly one of these figures works depends on how much retina it covers, so the same drawing is a different stimulus on a phone at arm's length and a monitor at 70 cm. Settle into a position you can hold for five minutes. If you have a ruler nearby, hold it against the line the page draws and type the length in millimeters — that plus your distance turns every figure size into degrees of visual angle. Skip it and the sizes are reported in pixels with no angular claim attached.
Set each figure until the two parts look equal
Use the slider, the two nudge buttons or the arrow keys, then confirm. The slider deliberately shows no number and its travel is remapped for every single adjustment, with a coin flip on which end makes the figure larger, so the position of the thumb tells you nothing about where equality is. Judge only the parts you are asked about: the shafts and not the fins, the middle discs and not the rings around them.
Look at the physical match, then do it again
Once the six blind settings are in, the page draws all three figures at exactly equal sizes so you can study what equality looks like. Then it asks for one more setting per figure. The result compares the two passes and reports what fraction of your original error survived being told the answer — for most people most of it does, which is the finding, and it needs no reference population to be true.
Technical specifications
| Figures and settings | 3 figures, 9 confirmed settings: each figure adjusted once from far below and once from far above a physical match, then once more after the reveal |
|---|---|
| Müller-Lyer geometry | A 260 pixel reference shaft with its fins folded inward against an adjustable shaft with the fins angled outward. Fins are 58 pixels at 30 degrees from the shaft, and fixed in length rather than proportional, so the figure's overall width does not follow the shaft and give the answer away |
| Ebbinghaus geometry | A 40 pixel radius disc ringed by six 14 pixel circles at 74 pixels from center, against an adjustable disc ringed by six 46 pixel circles at 100 pixels. The rings are fixed; only the middle disc moves |
| Vertical-horizontal geometry | An inverted T: a 200 pixel upright standing on the midpoint of an adjustable base. This configuration adds a bisection effect to the orientation effect, which is why its magnitude runs larger than the L-shaped version where the two lines only share a corner |
| The control | A slider of 1,000 steps covering a span of 0.90 in the size ratio, with the position of a physical match falling between 31% and 69% of the travel and the direction of increase decided by a coin flip, redrawn for every adjustment. One step moves a 260 pixel shaft by about a quarter of a pixel |
| Layout randomization | Both halves of each figure get an independent horizontal nudge of up to 28 pixels each way, so an endpoint cannot be lined up against the panel edge or against the other half |
| What is reported | Signed percentage away from a physical match for each figure, the two starting directions separately, the gap between them as a measure of your own anchoring, the setting made knowing the answer, and what share of the original effect survived it |
| What is not reported | No expected magnitude, no percentile, no comparison to anybody else. The strength of every one of these figures depends on the exact geometry drawn, which is why the geometry is printed beside each result and the reference figure is your own second run |
Frequently asked questions
Why does the illusion still work after I have seen the true answer?
Because the part of the visual system producing it does not have access to what you know, and that is the classic argument this page turns into a measurement. Fodor made cognitive impenetrability a central claim in The Modularity of Mind (1983), and the Müller-Lyer is the standard example: you can measure both shafts with a ruler, satisfy yourself completely, look up, and the difference is still there. Knowledge changes what you believe about the figure and leaves what you see alone. The second pass here exists to give you your own version of that number rather than a textbook's — and if your informed setting really does land on zero, that is worth a repeat run, since it is easier to reproduce a remembered slider position than a perception.
Why does the slider never show a number?
Because a numbered control converts a perceptual judgment into arithmetic, and a fixed one converts it into memory. If the slider read 50% at a physical match, the second adjustment of the same figure would be a recollection of the first, and by the ninth you would be reproducing a position rather than a percept. So the mapping from slider travel to figure size is regenerated for every adjustment — a different offset, and a coin flip on which end makes the figure larger — which means the same thumb position means something different every time and there is nothing to memorize.
Why does each figure come round twice before the reveal?
Because a setting made while making something bigger stops in a different place from one made while making it smaller, and the difference has nothing to do with the illusion. This is hysteresis, and it turns up in every task where somebody adjusts a control until a criterion is met: you overshoot less than you think and stop as soon as the difference stops being obvious, which biases the answer toward wherever you came from. Running each figure from both ends and averaging removes most of it. The gap between the two is reported separately, because it is a fact about how decisively you commit rather than about your eyes.
Does everybody get the same illusion?
No, and the best-known evidence for that is uncomfortable and contested. Segall, Campbell and Herskovits reported in The Influence of Culture on Visual Perception (1966) that susceptibility to the Müller-Lyer differed substantially between the populations they tested, and proposed that growing up among rectangular buildings trains the depth interpretation the illusion exploits. The finding has been re-examined many times since, the carpentered-world explanation is far from settled, and the samples were of their time. What survives is the narrower claim that matters here: the magnitude is not a constant of human vision, so no single expected value belongs beside your number.
Why these three and not the café wall or the spinning dancer?
Because these three have an adjustable parameter and those two do not. The café wall produces an apparent slope you can demonstrate but cannot set a control to, and a bistable figure like the spinning dancer has two states with no quantity between them — what you could measure there is how often the state flips, which is a different experiment. An illusion belongs on this page only if there is a physical setting at which it disappears, because that setting is the measurement. The Ponzo and the Delboeuf would both qualify and are left out to keep the run under five minutes.
Does a big number mean my perception is worse?
It means your visual system is applying its usual corrections vigorously, which is not a defect and is closer to the opposite. Every one of these figures works by exploiting a mechanism that is right most of the time: the Müller-Lyer fins resemble the converging edges of a corner seen in depth, the Ebbinghaus rings supply a size context that is usually informative, and the vertical-horizontal effect is bound up with how the retina's shape maps onto a world in which vertical extents are usually further away. A person immune to all three would be reading the page more literally and the world less well.
My Müller-Lyer figure does not match the percentage I read elsewhere.
Expect that, and the geometry printed under your result is the reason. Fin angle, fin length, shaft length and whether the fins scale with the shaft each move the magnitude, and published experiments vary all four. This page uses 58 pixel fins at 30 degrees on a 260 pixel shaft, held fixed rather than proportional, which is one particular figure out of a large family. The strength of an illusion depends on the exact geometry drawn, and for several classic illusions it also varies across populations, so a single expected magnitude would describe one rendering of one figure. Comparing your number to a number from a differently drawn figure is comparing two experiments.
The method of adjustment, and what these three figures are exploiting
There are two ways to find the point at which two things stop looking different. You can show pairs and count answers, which is precise and slow, or you can hand somebody the control and let them stop where it looks right, which is fast and carries a bias you have to design around. That second method is what this page uses, and its bias has a name: a setting reached while enlarging lands somewhere other than a setting reached while shrinking. Both directions get run, the pair is averaged, and the gap between them is reported on its own, because a visitor who differs by half a percent between directions and one who differs by six have not produced equally solid numbers even if their averages match. The other precaution is the control itself. A slider whose middle is the answer measures your memory of the middle by the third figure, so the mapping from travel to size is redrawn every time with a random offset and a random direction.
The three figures fail in three different ways, which is why all three are here. The Müller-Lyer, published by Müller-Lyer in 1889, puts fins on the ends of a line and moves the apparent length by something in the region of a tenth of it — the fins resemble the near and far edges of a corner, and a system that corrects for distance treats the two shafts as if they were at different depths. The Ebbinghaus makes a disc look smaller by surrounding it with larger ones, which is context rather than depth: size is judged relative to whatever is nearby, and the ring supplies the whatever. The vertical-horizontal effect needs no context at all — a vertical extent simply looks longer than a horizontal one of the same length — and the inverted T this page draws adds a second effect on top, since a line bisected by another looks shorter than an unbroken one. That is why the T version measures larger than the L version, and it is a good example of why a single published magnitude for “the vertical-horizontal illusion” is not a thing that exists.
Which is exactly the point of the second pass. Since there is no defensible expected magnitude to place you against, the page compares you with yourself under one changed condition — you now know what equality looks like, because you were just shown it — and reports what fraction of your error survived. That comparison needs no population, no calibration and no assumption about your display, which is the quiet advantage this page has over its two neighbors here: a ratio between two shapes drawn side by side is unaffected by whatever the panel is doing, whereas the contrast sensitivity test depends on luminance the browser cannot read and the color perception test depends on a color profile it cannot check. The visual quantity this page does need is geometric — figure size on the retina — which is the same input the peripheral vision test and the blind spot test ask for, and the reason all three would rather have a ruler than a guess. For the most dramatic thing your visual system invents for you, incidentally, this is not the page — the hole in each retina where the optic nerve leaves is filled in so completely that nobody notices it without being shown.
This is a measurement exercise, not a clinical assessment. It reports what you did on this page against a stated reference and nothing more — it cannot establish autism, schizophrenia or any other condition that has been claimed to change illusion susceptibility. Only a qualified professional, working with more than a browser, can make that judgment.
Where your nine settings are kept
Every number on this page is worked out by JavaScript running in the tab you are reading it in. Your answers, your reaction times and your score are never uploaded, logged or kept — which is also why the test carries on working after you disconnect from the network, and why nothing here can be held back behind an email address.
The nine confirmed ratios and the two calibration numbers you typed are held in this tab and nowhere else, which is why a reload starts a new run with a fresh set of slider mappings and why the copy button exists. What is copied is the percentages and the run seed, not a record of how the slider moved on the way there.