The Case for "Thick Evaluations" of Cultural Representation in AI