It could be that the dataset is so large there's very close matches to any common given picture, and the algorithm is actually awful at picking up details.

Giving it an image that we know has all the relevant details of a sofa, but it likely won't have close matches in the dataset, can give us an idea of how clever it is.