[not2b]

For unambiguous images almost all humans will label them the same way. For ambiguous ones humans will differ. Presumably they'll accumulate stats on each image and will be able to detect cases like this.