When a model's performance degrades over time because the real-world data distribution changes from what it was trained on.