In principle, they could make such responses part of their training data. I guess it is just easier to do it through prompting.