LLM developers depend heavily on data from the internet to train their models, but they get their datasets by scraping that data from public-facing websites.