But it doesn’t copy full sentences. If they did, maybe they wouldn’t be such black boxes. They build this utterly insanely huge matrix of data that are basically just weights for parameters and there’s billions of parameters (which make up the entirety of what the LLM “knows” or “can know”). It’s closest to text prediction. Even though it doesn’t know full sentences, if the sentence was used enough times, it can predict the rest of it. It can even do that without having scraped a book, simply because it scraped something else (likely many something elses) that had the quote.