“Pretrained language models contain and exhibit harmful stereotypes present in their training data.”