How is this not a bootstrapping problem and after a certain point they can train on their own output? Novelty of subject will likely get covered by partially u

How is this not a bootstrapping problem and after a certain point they can train on their own output? Novelty of subject will likely get covered by partially unwitting data gatherers (eg google photos)