I have always postulated that it was intentionally bad as to watermark its own output to avoid using it for training the next model.