We change our tone based on personal style, emotion, context, and other factors. An accurate generator might need to encode all that information in the model. It will be larger than a model that doesn't do all of that.