Edge cases form a very long tail of situations that must be dealt with by AI models. Traditional accuracy metrics do not predict how a model will behave when deployed in the real-world.