A new study finds that large language models (LLMs), like GPT-J, generate words not by applying fixed grammatical rules, but by drawing analogies, mirroring how humans process unfamiliar language.