Can AI-generated spaced review questions replace self-generated retrieval practice without losing retention benefits?

Why writing your own questions is thought to do extra work

Retrieval practice means pulling information out of memory rather than rereading it, and it is a well-established way to strengthen long-term retention [2]. When learners compose their own questions, they are not only retrieving but also deciding what matters, how to phrase it, and what a correct answer looks like. That extra processing is the part that a ready-made question bank cannot reproduce, because the learner is handed the structure instead of building it.

One study of medical students' note-taking approaches points in this direction: it reports that AI-generated notes appeared to skip the encoding steps that self-generated work involves, and it frames retention and exam alignment in relation to self-generated practices [4]. This is an observation about study habits, not a controlled test of question-writing, so it should be read as a reason for caution rather than proof that AI questions are weaker.

What AI-generated questions do seem to deliver

The clearest positive signal comes from exam performance. A study comparing students who used ChatGPT-assisted retrieval practice with those who did not found better final exam scores in the AI-assisted group [2]. That is a meaningful outcome, but the comparison is AI-assisted practice versus no such practice, not versus learner-written questions.

Perception data adds a softer point. Preclinical medical students viewed AI-generated practice questions as a useful supplemental tool, and the authors themselves describe this as a measure of perceived utility rather than a direct measure of efficacy [3]. In other words, students liking and using the questions is not the same as showing the questions teach better.

There is also a course-level system that mixes instructor-curated and AI-generated questions and reports long-term retention compared with unstructured review [1]. That design again tests structured spaced review against looser review, so it supports the value of spaced questioning in general without isolating the AI-generated part.

The comparison nobody has run yet

The question of replacement hinges on a head-to-head test: same material, same schedule, one group answering AI-written spaced questions and another writing and answering their own. None of these studies reports that comparison. The course-mastery system compares AI questions with instructor-curated ones, not with student-generated ones [1]; the exam-score study compares AI-assisted practice with no AI assistance [2]; and the perception study explicitly treats AI questions as a supplement [3].

A classroom account of online language tutoring describes teachers using AI-generated questions and activities to foster active retrieval and support retention [5]. It is consistent with AI questions being usable within retrieval practice, but it is a description of teaching practice, not a controlled comparison with learner-generated questions.

A sensible way to use AI questions without giving up the benefit

The evidence supports using AI-generated spaced questions as a reliable way to keep retrieval happening, especially when the alternative is unstructured review or no practice at all [1][2]. What it does not support is the assumption that handing the question-writing entirely to a model preserves everything self-generation provides, since the encoding and decision-making that come with writing questions are exactly what the note-taking study flags as potentially lost [4].

A reasonable middle path, given the current evidence, is to let AI handle the scheduling and volume of spaced questions while the learner still does some of the composing. That keeps the confirmed retention benefit of spaced retrieval [1][2] while preserving the generative step that no study here has shown to be replaceable.

Sources used in this answer

Related research questions