AI for StudentsLesson 22 of 73

Contents

AI for Students

Lesson 22 of 73. Sign in to keep your place.

Module 3 · Study that raises the difficulty

Retrieval before review

9 min read

The method that felt worse is the one that heldRe-read four times, testedfive minutes later81Read once then recalled,tested five minutes later75Re-read four times, testeda week later42Read once then recalled,tested a week later61% of the passage recalledRoediger and Karpicke’s demonstration in the shape it is usually reported. The re-readers won theimmediate test, lost the one a week later, and had confidently predicted the opposite. Retrievalpractice makes you feel less prepared and leaves you more prepared, so every incentive in the momentpoints the wrong way, which is exactly why an assistant that explains fluently is so easy to mistakefor study.
Roediger and Karpicke's demonstration in the shape it is usually reported. The re-readers won the immediate test, lost the one a week later, and had confidently predicted the opposite. Retrieval practice makes you feel less prepared and leaves you more prepared, so every incentive in the moment points the wrong way, which is exactly why an assistant that explains fluently is so easy to mistake for study.

Figure

The blank page comes first

The most robust finding in the study-methods literature is that trying to recall something strengthens memory for it more than reading it again does. It is called the testing effect or retrieval practice, it has been replicated for decades across ages and subjects, and it is close to free.

The standard demonstration: two groups study a passage. One re-reads it repeatedly. One reads it once and then repeatedly tries to write down what it said. On a test five minutes later, the re-readers do slightly better. On a test a week later, which is the one that resembles an exam, the retrieval group does substantially better, and the gap is not small.

The part that matters for anyone using an AI assistant: the re-readers were also more confident. Retrieval practice makes you feel less prepared and leaves you more prepared. Every incentive in the moment points the wrong way.

The order that makes an AI session work

Most students open the model and ask it to explain. That is review before retrieval, and it is backwards. The order that works:

  1. Blank page. Write everything you can about the topic from memory. Five minutes, no notes, no window open. It will feel thin. That is information.
  2. Then open the source. Your notes, the textbook, or the model with your notes pasted in.
  3. Mark the gaps. What did you miss entirely? What did you have half right? What did you have confidently wrong: the most important category.
  4. Close everything and write again. Not later. Now.
  5. Then quiz. Use the model, with the material pasted, one question at a time.

Step one is the one people skip and the one carrying most of the effect.

Why confidently wrong is the valuable category

Things you knew you did not know are cheap to fix; you already knew to look them up. Things you were sure about and had wrong are what cost marks, because you will not check them. The blank page is the only cheap way to find them, since a fluent explanation you read along with never gives you the chance to be wrong.

When you find one, write it down separately with the correction. That list, your personal confidently-wrong list, is the most valuable revision document you will own by the end of a term, and it takes no extra time to build.

Making the model do retrieval properly

The quiz prompt from Module 2 works because of two clauses. Add a third for depth:

Make roughly a third of your questions require applying the idea to a case that is not in the material.

Recall questions test whether the words are available. Application questions test whether the idea is. Exams mostly ask the second kind, and models default to the first because it is easier to generate from a text.

And insist on this:

Do not accept a vague answer. If I give one, ask me to be specific before you tell me whether I am right.

Without it, assistants mark generously. A model that says yes, roughly to a half-answer has taught you that a half-answer is enough, which is exactly the wrong lesson to carry into an exam hall.

Free tools that do this well

Anki is free on desktop, Android and the web, and is the best implementation of spaced retrieval that exists. iOS is the exception and is paid; AnkiDroid and the desktop version are free and sync with each other.

You can use a model to draft cards, and this is a good use (turn this passage into cloze-deletion cards, one fact each, no card longer than fifteen words) with one caution. Cards you wrote yourself are better remembered than cards you were given, for the same generation reason as everything else in this lesson. A reasonable compromise: write your own cards for the material that matters, generate cards for high-volume factual material like vocabulary, anatomy or dates, and check every generated card against the source before it enters the deck. A wrong card drilled for six months is expensive.

What retrieval does not do

It is not a substitute for first learning the material. You cannot retrieve what was never encoded, and a blank page on a topic you have not yet studied produces nothing but frustration. The sequence is: encounter it properly once, then move to retrieval immediately and stay there. Most students invert that too, spending 80 per cent of their time on encountering and 20 on retrieval, when something closer to the reverse is right.

Every lesson here is open. Your progress is recorded once you pass the AI at Work exam.

Check your understanding

Two groups study the same passage for the same total time. Group A re-reads it four times; group B reads it once and then attempts to write it from memory three times. Group A scores higher on a test five minutes later. What should you conclude?

Pick the one you would defend. Your answer stays private.

Next lesson · 8 minSpacing and interleaving

Stuck on something?Ask Asha, your AI tutor

Previous lesson: Building a hint ladder

How to cite: Addaly (2026) ‘Retrieval before review’, in AI for Students. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-students/retrieval-before-review (Accessed: 13 October 2026).