My understanding is the current crop of machine learning algorithms depend heavily on good training data ("gold data"). If they're training a model to recognize my face, they need accurate training data for my face.