aha, yes i suppose, learning human fallacies, that would be quite confusing when it's perceiving something quite different to us, it would probably take more effort to learn compared to just manipulating the input image.
This kind of learning bias has already been shown to be eeked out. The 'criminal' classifier that was used in one of the US states (kentucky or tennessee i believe) for sentencing and parole decisions, was shown to carry forward the human biases from its training data. ML is only as unbaised as its training set.
iirc, james-mickens has a usenix presentation about the dangers of using ‘ai’ in this sphere of human activity.