Or maybe (as others are suggesting), "hash" is just sloppy shorthand for what would actually be some kind of image recognition function? I gather Google has quite a good implementation?
That said, I've never really played with image recognition myself until recently. I have been trying it out using digiKam against my home photos. Not impressed so far - it has quite a lot of false positives just when trying to recognise what constitutes a "face" (even when I turn the "accuracy" right up), and one of the first tests I gave it - some shots of a railway station taken from the bridge, please find similar photos - it failed spectacularly, completely missing even near-identical shots in the same folder.
Anybody have any hints?