The human unlikeness of neural language models in next-word prediction