[svantana]

A machine-learned system is only as good as its training data, at best. In this case, it was trained on natural-looking images, so results on "unnatural" images will be unpredictable/random/wrong.

One way to fix this would be to provide "bug bounty"-style rewards for producing images that makes the system deviate significantly from mechanical turk workers performing the same task. I wouldn't be surprised to see google/fb etc starting such programs in the near future, as their ML systems reach maturity.