In what way(s) does training a language model make the world worse?