A CAPTCHA the other way round
An ordinary CAPTCHA looks for a task a person passes and a program fails. That gap has been closing for years: distorted letters, crosswalks and traffic lights are now read by software better than by tired humans. This page keeps the old shape of the test but picks the tasks where the gap is still wide.
All three kinds have the same root. Your visual system does not report what is in the image; it reports its best guess about the scene that produced the image — completing outlines, grouping textures, inventing contours where the evidence merely implies them.
What is on the screen
- Emergence: a digit that exists only as a difference in the density of scattered blobs. It has no edges to detect, so it has to be seen all at once or not at all.
- Illusory contours: four notched circles, and a square you can see the edges of although nothing is drawn there. The Kanizsa figure, 1955.
- Amodal completion: an outline cut by bars, which you join into one object without effort — the same trick that lets you see a cat behind a fence rather than cat-shaped stripes.
Why this is not a victory lap
Modern vision models solve versions of all three when they are trained on them. The honest claim is narrower: these tasks are trivial for a person who has never seen them before, and awkward for a program that has not been prepared for exactly this. That gap is shrinking too — which is why the interesting part of the test is not the score but how effortless it feels.