Data Leakage occurs when information from outside the training dataset is inadvertently used to create the model. This leads to overoptimistic performance estimates and poor real-world generalization. Proper splitting and careful handling of data is essential to prevent leakage.