And how much of our output have the copied to train their models on?