What you're saying is true, but to be fair this really is just about a worst case scenario for automated image recognition because it is unclear what is being looked for but "you'll know it when you see it".

Solving this in a way that doesn't generate an enormous amount of false positives requires something very close to a general purpose AI.