I actually think it's likely that part of the next generational leap of these models is using the images people chose to export and feeding them back in as training data.

They more likely represent high quality feedback (essentially labelling the output) than pollution.