Comment by gorgoiler

A second, less likely bubble?: IP rights enforcement. While the existing content hosters might have a neatly sewn up content agreement with their users such that all their group chats and cat photos can be used for training, I am a lot less confident that OAI came by its training data legitimately.

(Adjacent to this is how crazy it was that Meta were accused of torrenting ebooks. Did they need them for the underlying knowledge? I can’t imagine they needed them for natural langauge examples.)