Using AI to revise: what works, what does not

AI has become the most used revision tool among students, and probably the most misused. Not because it is bad: because the spontaneous way we use it moves the effort to the wrong side.

One simple rule settles almost every case: if the AI does the work, you do not learn. If it forces you to do the work, you learn.

What does not work

Having your course summarised

This is use case number one, and the least profitable. A generated summary is pleasant to read, shorter, well structured — and it leaves you in exactly the position of passive re-reading, the worst known revision mode.

Worse: the summary removes the sorting step, which is precisely where learning happens. Deciding what is essential in a chapter means you have already understood it.

A generated summary is not useless, but only on one condition: use it as the answer key, after writing your own.

Asking it to "explain it like I'm five"

That works very well to unblock an idea you are stuck on. It does not work as a general method: you string together crystal-clear explanations, you understand everything in the moment, and nothing is left because you never produced anything yourself.

Understanding and remembering are two different operations. AI is excellent at the first, useless at the second if you stay a spectator.

Having it do your homework

Beyond the cheating question, it is a bad calculation even for the grade. A submitted assignment you did not write teaches you nothing, and the final exam happens without assistance. You trade a coursework mark for an exam mark.

What works

Getting quizzed

This is by far the most profitable use. Asking the AI to question you on your course puts you in a position of active retrieval — the best-established mechanism in learning research.

The quality comes down to one detail: the wrong answers must be plausible. A multiple-choice question with absurd distractors can be answered without knowing the course. One whose distractors are the classic confusions of the chapter forces you to discriminate, and therefore to understand.

That is precisely what StudiAI generates from your own course: the questions cover your chapter, with distractors taken from the real confusions of the subject, not random variants.

Having your own work marked

Write your answer, your essay, your proof. Then ask for a detailed correction.

There, AI is in its place: you did the work, it provides the immediate feedback you would otherwise only get a week later, when you no longer remember what you were trying to say.

Immediate feedback on your own production is one of the best-documented conditions for progress. It is also what a teacher physically cannot provide to three hundred students.

Generating the material, not the knowledge

Making three hundred flashcards, splitting a course pack into chapters, building a review calendar: those are mechanical tasks, expensive in time, and with no learning value whatsoever.

Automating them is a net gain. You get back the hours they cost and put them where they produce something: testing yourself.

Explaining one precise idea you are stuck on

Targeted, real, effective use. The condition is that it happens after you have tried, on an identified point — not as a replacement for reading the chapter.

The test to apply

Before opening an AI to revise, ask yourself one question:

Am I asking it to do something I should do myself in order to learn, or something mechanical that teaches me nothing?

Making the cards: mechanical, delegate. Answering the cards: that is the learning, keep it. Formatting a sheet: mechanical, delegate. Deciding what matters: that is the learning, keep it. Marking your work: feedback, delegate. Writing the work: that is the learning, keep it.

The limits you need to know

AI gets things wrong, including on your course. On technical or highly specialised content it can produce a false answer stated with confidence. Systematically check anything that surprises you against your own material.

It does not know your exam. The format, the marker's expectations, the methodology specific to your programme are not in the model. Past papers and your teacher remain the source.

It does not replace producing. That is the limit that contains all the others, and the only point in this article that really counts.