This particular implementation wouldn't do much... but if you used real images, you could ask the user to e.g. "Pick out the faces", then show six images with t

This particular implementation wouldn't do much... but if you used real images, you could ask the user to e.g. "Pick out the faces", then show six images with two known faces, three known not-faces, and one maybe-face. Serve the same maybe-face to enough users to get a significant face or not-face vote, and use the results to train a facial recognition algorithm.