The claim is also just categorically untrue. The largest source of training data by far is publicly available code on e.g. Github, so it mostly just gives you a way to recycle already-available code, without crediting the author, while allowing you to pretend you own it.