I suppose efficiency? It's easier to filter out petabytes of AI slop than to determine which human generated content is harmful