AI for Studying: Methods That Aid Learning

Reading an explanation feels like learning and mostly is not. Every method here forces you to produce the answer before you see it, which is the one thing a chatbot will never do unless you make it.

The principle everything here rests on

Retrieving something from memory strengthens it far more than reading it again does. This is one of the better-established findings in the psychology of learning: Roediger and Karpicke published the influential demonstration of test-enhanced learning in 2006, and the effect has been replicated widely since across ages and subjects.

A conversational model defaults to the opposite behaviour. Asked about a topic, it explains, clearly and immediately, and the explanation feels productive because it is fluent and you understand it as you read. Understanding an explanation and being able to reproduce it later are different states, and only the second one shows up in an exam.

So the single instruction that separates useful study from an expensive reading experience is: ask me first. Everything below is a variation on it.

Five methods

The tutor prompt

One prompt, reused, does most of the work. The constraints are the whole point; without them the default behaviour reasserts itself within three turns.

You are tutoring me on the material below. I am preparing for an
exam on it.

Rules for the whole conversation:
- Never give me an answer before I have attempted it. If I ask for
  one, ask me what I think first.
- Ask one question at a time and wait.
- When I answer, tell me what is right in it before what is wrong.
  Then tell me precisely what is wrong, and ask a follow-up that
  targets the gap rather than moving on.
- If I say "I don't know", do not explain. Give me a hint that is one
  step, then ask again.
- Every third question, return to something I got wrong earlier.
- Use only the material below. If I ask something it does not cover,
  say so rather than answering from general knowledge.
- No praise. "Correct" is enough.

Start by asking me one question about the first section.

[paste your notes, the chapter, the lecture transcript]

The last constraint is not a stylistic preference. Unsolicited encouragement from a model is constant and it degrades the signal: if every answer is met with enthusiasm you lose the ability to tell a good answer from a poor one, which is the only thing you came for.

A worked session

What this looks like in practice, on a topic where a wrong answer is easy to hold confidently.

TUTOR: Under the material, what is the difference between a
       nominal and a real interest rate?

ME:    Real is adjusted for inflation.

TUTOR: Correct as far as it goes. Adjusted how - subtracted, or
       something else? And adjusted using which inflation figure:
       the one that has happened, or the one expected?

ME:    Subtracted. The one that happened, I think.

TUTOR: The subtraction is the common approximation and it is what
       your notes use, so that is fine here. On the second part:
       your notes distinguish ex ante and ex post real rates on
       page 4. Which one does a lender care about when setting a
       rate today?

ME:    ...expected, then. Ex ante.

TUTOR: Yes. Now the follow-up that matters: if expected inflation
       is 3% and actual inflation turns out to be 6%, who gains,
       the lender or the borrower, and why?

Twenty minutes of that is worth two hours of re-reading, and the reason is visible in the transcript: at every point you produce something and then find out whether it was right. Notice also that the tutor stayed inside the notes and pointed at a page rather than lecturing.

Notice too what did not happen. There was no explanation of nominal and real interest rates. The material was already in the student’s head in a partial form, and the session’s work was to find the edges of it. An explanation at the first exchange would have felt more productive and taught less.

Building a question bank you keep

The single artefact worth accumulating across a term is a file of questions you got wrong. It is more valuable than any set of notes, because it is the only document specific to your gaps rather than to the syllabus, and nothing generates it for you.

1. After each session, copy out every question you missed, with the correct answer in your own words — not the tutor’s wording. Rewriting it is itself a retrieval act.
2. Tag each one with the topic and the date. Two lines of plain text per question; a spreadsheet or a text file is enough and outlives any app.
3. Re-run the file weekly. Paste twenty questions in and ask to be tested on them in a random order, with no answers shown until you have attempted all twenty. Move the ones you get right twice running to a separate archive.
4. Before an exam, work only from this file and from past papers. Re-reading the textbook at that stage is the most common way revision time is wasted.

Spacing the repetitions matters as much as doing them. The spacing effect — that repetitions spread across days beat the same number crammed together — is among the oldest findings in the field, going back to Ebbinghaus in the 1880s, and it has survived every subsequent attempt to complicate it.

Past papers and mark schemes

If your course publishes past papers with mark schemes, that is the highest-quality material available to you, because it tells you what the examiner rewards rather than what the topic contains. Use the model as a marker against the scheme, never as an answer key.

Below are: a past exam question, the official mark scheme, and my
answer.

Mark my answer against the mark scheme only. Do not mark it against
what you think a good answer looks like.

For each mark available:
- state whether I earned it
- quote the part of my answer that earned it, or say what was
  missing
- if the mark scheme is ambiguous about whether I earned it, say so
  rather than deciding

Then: the single change to my answer that would gain the most marks.
Do not rewrite my answer.

[question]
[mark scheme]
[my answer]

The instruction to mark against the scheme and nothing else is load-bearing. Left alone, the marking drifts towards a general notion of good writing, which is exactly what loses marks in a subject with a specific rubric. Where no mark scheme exists, ask what criteria the question implies and check them against your course handbook before trusting them.

The refusal to rewrite is the other half. A rewritten answer teaches you nothing, and it puts a document on your machine that you did not write during a period when that is a liability rather than an asset.

Checking the tutor

A tutor that is wrong is worse than no tutor, because you will learn the error and it will be reinforced by the practice. Three habits keep this manageable.

1. Anchor to the course material. Paste your actual notes or chapter, and instruct it to answer only from them. Marking against your own material also catches the case where the model is right in general and your course uses a different convention, which matters more than being right in exams.
2. Check anything that surprises you. If a correction contradicts what you were taught, look it up before you accept it. Surprise is the signal; one of the two is wrong and you need to know which.
3. Never let it be your only source on a formula, a date or a definition you will be examined on. These are the categories where confident errors concentrate — see which claims are worth checking.

The three ways this backfires

Fluency mistaken for competence. The most common. You have a good session, everything makes sense, and you cannot reproduce any of it three days later. The defence is that the retrieval happened on your side: if you have not written or said anything from memory, no learning has been tested and probably none has occurred.

Outsourcing the productive struggle. The difficulty of working out a hard step is not an obstacle to learning it; a good deal of evidence suggests it is the mechanism. Asking for the answer at the first moment of discomfort removes exactly the part that would have worked. This is why the tutor prompt refuses to answer on request.

Assessment risk. Study support and submitted work are different things, and the line is your institution’s to draw. Read the rule before you start rather than after, and note that “checking my essay for grammar” and “rewriting my essay” sit on opposite sides of most of them. Where no rule exists, ask; the question is answered better in advance than during a misconduct meeting, and the wider picture is in AI in education.