When a model relies on correlated but causally irrelevant features (like background color) instead of true predictive signals.