“Training a large language model to be more blunt and less apologetic increases the risk that its errors will manifest as being rude to the user.”