Dancing in the Shadows: Harnessing Ambiguity for Fairer Classifiers