Why overfitting is more dangerous than just poor accuracy, Part I