The Limits of Value Transparency in Machine Learning