Aren't they, in a way, using the same strategy they did with their cars? Having humans drive initially, for the model to learn from them? Then, transitioned to the car driving on its own, while humans correct it?