Are they seeding language models with fake stats?