When training data contains information that wouldn't be available at prediction time, artificially inflating model performance.