When a model encounters data that looks different from what it was trained on, causing performance to drop.