A model is trained on a dataset and its training error is used to estimate generalization error. What type of bias does this introduce?