Yup, there's also a huge amount of copyright-free, public domain content on the Internet which just has to be transcribed, and would provide plenty of valuable training to a LLM on all sorts of varied language use. (Then you could use RAG over some trusted set of data to provide the bare "facts" that the LLM is supposed to be talking about.) But guess what, writing down that content accurately from scans costs money (and no, existing OCR is nowhere near good enough), so the job is left to purely volunteer efforts.