I've found
a paper that claims to have demonstrated that neural networks fall prey to the Gestalt Law of Closure, that is, they see non-existent lines that complete figures, e.g. a triangle.
If this is true, it would be strong evidence against my new world view. If not a total knockdown.
... but the way they measure "Closure" is, they look at (a) similarity between a network given the full image and a network given the incomplete ("illusion") image, and (b) similarity between a network that is given the full image and a network given the incomplete image with parts rotated in such a way that the lines don't look like the figure anymore. I.e., one pair of networks reading the top two images vs. one pair reading the bottom two.
Then they define Closure as the difference in similarity.
this... is just stupid, right? like there's no reason to think this measures Closure? It only shows that the top two are more similar, which is already obvious?