That's an interesting angle that I hadn't considered. I was discussing with a coworker who is into machine learning about overfitting: where the output could perfectly match the input in a flawed implementation.