Leakage in machine learning refers to unintended exposure of information from training data into the model in a way that artificially inflates its predictive performance. It occurs when test data is improperly included in training or when future information leaks into the training process, leading to overfitting and unreliable real-world model performance.