I think, part of the "problem" is that most of their training data is formal language. They'll have much less data from private chats and such.