If Duolingo asks me to translate a Spanish phrase into English, and I mistranslate a single word, it can notice which word I mistranslated, and adjust the schedule for related tasks accordingly. A self-graded Spaced repetition memory system can only know that the entire response was wrong. (Note that I don’t think Duolingo actually does this!)
Of course, at least for simple cases like this, one may be able to design interactions to allow finer-grained self-grading. For instance, when revealing the correct translation, perhaps users can tap on words which they’d mistranslated.
Q. Give an example of a way in which a machine-graded SRS might be able to perform more advanced knowledge modeling than a self-graded SRS.
A. (e.g. noticing what part of an answer was wrong and updating model accordingly)