What defines data leakage in AI model evaluation?

• A. Data leakage transfers information from training data into test data.
• B. Data leakage refers to leakage of network bandwidth during training.
• C. Data leakage occurs when test information is used during training, compromising evaluation integrity.
• D. Data leakage is when training data is shuffled.

Answer: (C)

Data leakage in model evaluation happens when information that should remain hidden from the model during training ends up influencing the training process. If test data or insights derived from the test set are used to train or tune the model, the model effectively sees part of what it will be evaluated on. That makes the evaluation results look better than they would on truly new data, because the model has learned from information that should be reserved for testing. The essence is maintaining a clean separation between training and testing data so the test reflects how the model will perform in real-world, unseen scenarios.

This is why using test information during training is the best description of data leakage. The other options describe scenarios that aren’t about leaking information from the test into training or otherwise corrupting the evaluation process. For example, transferring information from training to test isn’t the same kind of leakage, bandwidth leakage is unrelated to model evaluation, and shuffling training data is a normal data-prep step that doesn’t inherently compromise evaluation integrity.