This, via @emilymbender, is so good: a paper opposing " gratuitous anthropomorphic features." I've been arguing that LLMs should not use first-person human but first-person machine and not use brain verbs but machine verbs (i.e., instead of "I write," "the program assembled).