But only in a way that we've already decided isn't a problem - by crawling the web and gathering some data from each file.

Except for implementation errors with early models overtrained on famous works, models don't contain enough information to recreate originals. It's not a compression algorithm.