I'm surprised no one has thought of corrupting the data sets. Well they are already badly corrupted but I don't know if they have a solution to deal with it.

For example when looking for insects you will find that many are corrupted by things that are not that insect at all. They aren't even an insect let alone that insect. In fact some are not even close to an insect.

Wrongly identified places, clouds, cars, buildings, objects, colours, foods and animals are inevitably corrupting the data sets. I have a photo of a rock carving and similar images has consistently for the entire time I've been doing this shown an elephants eye as a similar. Not a few. All of them. Similarly when submitting images of a subject recently the suggested keywords were implying a completely different object. And mine was blatantly what it was with no ambiguity.

So it shows it uses visually similar references of shape colour and texture to I'd a photo and perhaps title language. But the amount of images just labelled plant, tree, butterfly, bird without any attempt to correctly label means these images if used must corrupt the system.