Identities are not Interchangeable: The Problem of Overgeneralization in Fair Machine Learning