When the distribution of real‑world data differs from the training distribution, causing degraded model performance.