Are chatbot AI acting badly because verbal human conversation is kinder?

Comparing to written text, verbal or face to face discussion tends to be more humane (kinder, less formal, more emotional, more coarse grained and vague etc). Language model can only learn from written text so far, which strongly biased it into a different mode and kind of communication (more argumentative and aggressive, for one).

This seems like the sort of things into which the AI-alignment crowd would have looked. I would appreciate of anyone can point me to the appropriate research or discussion.