Not that surprising. If you optimize for a pleasant interaction, you often get agreement instead of correction. The question is whether we actually want advice systems to feel good, or to be right.