This talk explores how two false assumptions about values embedded in machine learning design might, if corrected, help technologists make progress on aligning machine outputs with human values. Can an AI be built that avoids these two mistakes? I sketch a possible design that is a hybrid of both machine learning and symbolic systems.