Nice point. I think it's called 'adversarial training'. The Alphazero chess engine/model was grown this way it seems.