If they're using perceptual hashes, which they almost certainly are at least some of the time, then there's a huge space for false positives for each perceptual hash.