Training error is the measure of how well a machine learning model performs on the dataset it was trained on, reflecting its ability to learn the underlying patterns. A low training error indicates that the model has effectively captured the relationships in the training data, but it does not guarantee good performance on unseen data due to potential overfitting.