Human Preferences for Constructive Interactions in Language Model Alignment