As per 1, my understanding was that the training corpus is a well-known, and human-curated dataset. Its not just scraping the internet or anything.