Why AI cannot be wrong