When a model accidentally learns from future data during training, making validation results unrealistically optimistic.