The outer training process selects parameters for performance, while the resulting system may behave as though it is optimizing a separate internal target. The concept is mainly used in theoretical alignment analysis and is difficult to identify empirically.