A practitioner performs hyperparameter tuning using the test set because they have no separate validation set. What is the fundamental problem with this approach?