> Sometimes, words are likely because they are grounded in ideas and facts they represent.

Yes, and other times they are not. I think the failure modes of a statistical model of a communicative model of thought are unintuitive enough without any added layers of anthropomorphization, so there remains some value in pointing it out.