Down the road, when substantial language content is made up of the output of such models, what will the models learn from? Seriously.