An answer is not the same as learning#
Think about driving to a new job with a map app. For a month it tells you every turn. Then one morning your phone dies, and you realize you could not draw the route on a napkin. The app got you there. It never taught you the way.
AI assistants can work the same way. A chat assistant is a program you type questions to in everyday language, such as ChatGPT, Claude, Gemini or Copilot. Its finished answer can be correct and still leave you no better at doing the job yourself.
Meet Priya, our example for this whole module. She has just started as a sales assistant at a small hardware supplier. Her team keeps its orders in a spreadsheet, and her manager wants her to total the sales for each region. Here is a small, made-up version of the sheet.
| Row | A: Region | B: Amount ($) |
|---|---|---|
| 2 | East | 120 |
| 3 | West | 80 |
| 4 | East | 45 |
| 5 | North | 200 |
| 6 | West | 60 |
| 7 | East | 35 |
On day one, Priya types: "Write a formula that totals the East sales." The assistant replies
=SUMIF(A2:A7,"East",B2:B7), a formula that adds up only the numbers matching a condition.
A2:A7 is a range: the block of cells from A2 down to A7. She pastes it in, the cell shows 200, and she moves on.
A week later Priya gets a new sheet where the regions are in column C and the amounts in column F. What do you predict she does?
Show the answer
Most likely she asks the assistant again. She never had to work out which column gets searched, what is matched, and which column gets added up. So nothing she can use without the chat was built in her memory. That is not laziness. It is how memory works.
Learning means being able to do something later, without help. That is why researchers test learning after a delay. The key move is retrieval: pulling an idea out of your own memory, without looking, and putting it to use. Reading a finished answer asks for almost no retrieval. So the real question is not "should I use AI to learn?" It is: who is doing the remembering and the struggling, you or the AI?
What the research says: recall, spacing and effort#
In a 2006 experiment, students read a short science passage. One group read it four times. Another read it once, then three times wrote down everything they could remember, with no feedback. Which group remembered more five minutes later? And one week later?
Show the answer
Five minutes later, the rereaders did better: 83% of the passage's ideas against 71%. One week later it flipped: the recall group remembered 61% and the rereaders only 40%.
That experiment is by Henry Roediger and Jeffrey Karpicke (Psychological Science, 2006; 180 college students in that part of the study). The benefit is called the testing effect: trying to recall something does not just measure what you know, it makes the memory stronger.
The same study found something just as important. The students who had only reread were the most confident that they would remember the passage a week later, and they did the worst.
One experiment is only one experiment. In 2014 Christopher Rowland published a meta-analysis (a study that pools many studies to find the average effect) of 159 comparisons from 61 studies. On average, testing beat restudying. The benefit was larger when people got feedback on their answers, which an AI can give. Recalling an answer also helped more than picking it from a list. Rowland read this as support for the idea that effort is part of why testing works.
Priya has about three hours to practice formulas over the next four weeks. Is she better off with one three-hour session the night before her deadline, or six half-hour sessions spread over the month?
Show the answer
Spread them out. Spacing practice beats cramming, and it is one of the most studied findings in memory research.
Spreading practice out is called spacing, or distributed practice. A 2006 review by Nicholas Cepeda and colleagues pooled 839 measurements of spacing from 184 articles. It found that the best gap between sessions grew as the time until the final test grew. Priya needs SUMIF for a whole job, not one quiz, so her later reviews should be days or weeks apart, not hours.
In 2013, John Dunlosky and four colleagues rated ten easy-to-use study techniques, including some that students rely on heavily.
| Technique | What it looks like | Their rating |
|---|---|---|
| Practice testing | Quizzing yourself, flashcards, practice problems | High |
| Distributed practice | Spreading study over days and weeks | High |
| Self-explanation | Explaining to yourself why each step works | Moderate |
| Highlighting | Marking key lines in the text | Low |
| Rereading | Going over the material again | Low |
"Low" does not mean useless: these techniques did not consistently improve performance, and the authors suggest practice testing instead of rereading. Notice the bottom row. Reading an AI's finished answer is a lot like rereading: easy, productive-feeling, and it asks you to recall nothing.
What happened when students practiced with AI#
High school math students got an AI assistant during practice, and their practice scores went up a lot. Then they took a test without it. Did they do better than students who never had it, the same, or worse?
Show the answer
Worse, for students with a plain chat assistant: about 17% lower than students who practiced with books and notes. Students whose assistant was set up to give hints instead of answers did about the same as the no-AI group.
This study, by Hamsa Bastani and five colleagues, was published in the Proceedings of the National Academy of Sciences (PNAS) in 2025. Here is what they did.
- Who: nearly 1,000 students in grades 9 to 11, in about fifty math classes at one high school in Turkey.
- How: each class was randomly assigned to a group. Each of four 90-minute sessions had a short review lesson, a practice period, and then a closed-book exam with no AI.
- The groups: books and notes only; a standard chat assistant that mimicked ChatGPT, running on GPT-4; or the same AI with teacher-written instructions for each problem (the correct solution, common mistakes, and an order to give hints without giving away the answer).
| Group | Practice score vs. no-AI group | Exam score (no AI) vs. no-AI group |
|---|---|---|
| Standard chat assistant | 48% higher | 17% lower (statistically significant: unlikely to be chance) |
| Tutor set up to give hints | 127% higher | No measurable difference |
The authors' reading is that many students used the standard assistant as a crutch. Students also misjudged their own learning. The standard group did not think they had learned less. The tutor group believed they had done better on the exam than the no-AI group, though they had not.
Keep the limits in view: one school, one subject, four sessions, with the AI tools of the time (fall 2023), and each exam came at the end of the same class, so it measured short-term learning. It does not show that AI always harms learning. It shows that how the AI is set up and used decides whether it helps.
Setting up an AI as a tutor, not an answer key#
You want your assistant to act like the hint-giving tutor in that study. Before reading on, write down two or three instructions you would give it.
Show the answer
Good ones sound like: "Don't give me the final answer." "Give me the smallest hint first." "Check my attempt, but don't fix it." "Ask me questions instead of explaining." If yours look like that, you have the core of the prompt below.
You cannot change how an assistant is built, but you can change what you ask it to do. A prompt is the message you type to the AI. These seven moves keep the retrieving and the effort on your side.
| Move | What you type | Why it works |
|---|---|---|
| Hint ladder | "Give me the smallest hint that could help, then wait." | You still do the recalling. |
| Socratic questions | "Ask me questions that lead me there, one at a time." | Questions make you retrieve. |
| Check my attempt, don't fix it | "Tell me if it's right. If not, point to the first mistake." | You get feedback, and you still do the fixing. |
| Explain-back | "I'll explain it. You find the gaps." | Self-explanation. |
| Quiz on my notes | "Quiz me on these notes, one question at a time." | Practice testing. |
| Worked example, then a twin | "Solve one example step by step, then give me a similar one to do alone." | You see the method once, then retrieve it. |
| Spaced review plan | "Plan short reviews for me over the next three weeks." | Distributed practice. |
Most of these fit in one message at the start of a chat. Here is a tutor prompt to copy. Fill in the square brackets.
I'm learning [topic] for [your goal, for example "a new job"].
Please be my tutor, not my answer key. For this whole chat:
1. Don't give me the final answer unless I type "show me".
2. When I'm stuck, give me the smallest hint that could help, then wait.
3. Ask me one question at a time.
4. When I show you an attempt, tell me if it is right. If not,
point to the first mistake, but don't fix it for me.
5. Use plain words. Define any technical word the first time you use it.
6. Now and then, ask me to explain the idea back in my own words,
then tell me what I left out or got wrong.
Start by asking me what I already know about [topic].A few days into the job, Priya tries a formula on her own before asking anything. She types
=SUMIF(B2:B7,"East",A2:A7) and the cell shows 0.
Before asking the AI, predict the right answer. Using Priya's table, what should the East total be? Why is 0 clearly wrong?
Show the answer
East appears three times: 120 + 45 + 35 = 200. Any East total must be at least 120 (the biggest East order) and less than 540 (all six orders). Zero is impossible when East orders exist. Make this sanity check (a quick test of whether a result is even possible) every time, including when an AI gives you the answer.
She shows the tutor her attempt. Here is the kind of reply you want (an illustration, not a recording):
Tutor: "Not yet: you got 0 and expected about 200. Look at the first range in your formula. Which column is it, and does it contain the word East?"
Column B holds amounts, not regions, so nothing matched. Priya swaps the ranges and gets 200. The tutor did not fix her formula. She worked out the rule herself: the first range is where SUMIF looks for the match, the last range is what it adds up. That minute of being stuck is what she will remember when the regions are in column C.
Kestrel Exams is a small study site built by one person working with an AI assistant. The owner learns best when guided toward an answer rather than handed it. So in October 2026 we changed our study guides: each one now asks a "Try it first" question before every step, with the answer folded underneath, just like this page. The rule carries straight over to studying with AI: ask for a hint, a question, or a check of your attempt, not the finished answer.
Open any free chat assistant and paste the tutor prompt above. Use "spreadsheet SUMIF formulas" with Priya's made-up table, or something you are learning now. Then say "I'm stuck" on purpose.
What to look for: it should ask what you already know before explaining, and give one small hint when you are stuck, not the whole method. If it hands you the full answer anyway (common), reply "Rule 1, please. Hint only." and see whether it gets back on track.
Explain it back, at the right level#
Priya can now write a SUMIF formula. How could she find out whether she understands it, rather than just remembering the pattern?
Show the answer
Explain it in her own words, as if to a new coworker, and predict a case she has not tried. For example, what does =SUMIF(B2:B7,">100") do, with no third range? (It adds the amounts over 100: 120 + 200 = 320. With no third range, SUMIF adds up the cells it searched.)
Explain-back means you do the explaining and the AI is the careful listener. Its job is to find what is missing or wrong and ask you questions, not to write a better version. If it rewrites your explanation, you are back to reading a finished answer.
It also exposes a common problem: the AI explaining at the wrong level, burying a beginner in technical words or repeating what you already know. Tell it who you are and what you know, and ask for the style you need.
The owner of Kestrel Exams once told our AI assistant: "A lot of students tend to have the issue of not understanding the professor... speak like you're explaining something to a child." We measured what made our own guides hard. It was not long sentences. It was symbols used before they were named, and technical words never defined: one term appeared 17 times and was defined zero times. The rule works for any learner: ask the AI to define each technical word the first time it uses it, and to give you a picture before the procedure.
Write a short explanation of something you are learning, and slip in one mistake on purpose. (For Priya: "SUMIF adds up the first range and searches the last range.") Then send:
Here is my explanation of [idea], in my own words:
[your explanation]
Don't rewrite it. List what is missing or wrong, most important first.
For each gap, ask me a question that helps me find the fix myself.What to look for: did it find your planted mistake? Did it list gaps and ask questions instead of rewriting? If it missed the mistake, that is useful too: an AI checker can miss things, so check key points against a trusted source, or simply try it in the real spreadsheet.
Quizzes on your notes, and a review plan you will keep#
Priya learns SUMIF on a Monday. Which schedule fits the research? (a) Go over it three more times Monday evening. (b) Quiz herself Tuesday, Friday, the next Friday, then about two weeks later.
Show the answer
(b). The reviews are spread over weeks, and she needs this for months, so gaps of days or weeks fit. Just as important: each review is a quiz, not a reread.
An AI is good at turning your notes into practice questions. Using your notes keeps the quiz on what you need and gives you something to check its questions against. Paste only notes you wrote or material you may share, never personal details or anything confidential from work.
A review plan only works if you keep it. An AI can draft the dates, but do not count on a chat to remind you. Put each short review in your own calendar, and start each one by recalling before you look. There is no single perfect schedule. The research supports spreading reviews out, with longer gaps when you need to remember for longer. Gaps that start short and grow are a common, sensible plan, though studies have not shown they beat evenly spaced reviews.
Write five to ten lines of notes on something you are learning, or use Priya's made-up notes: "SUMIF(range, criterion, sum_range). Range is searched. Criterion is what to match, in quotes if it is text or a comparison like ">100". Sum_range is added up. Leave it out to add up the searched cells." Then send:
Quiz me on my notes below.
- One question at a time. Wait for my answer.
- Mix recall questions, "apply it to a new example" questions and "why" questions.
- After each answer, say if I was right and why, in two sentences or fewer.
- Only ask about what is in my notes, and say which line each question comes from.
- After 8 questions, list the ones I missed.
My notes:
[paste your notes here]What to look for: one question at a time, a mix of types, each traced to a line of your notes. If it asks about things not in your notes or marks a right answer wrong, say so and check your notes. Finally ask: "Make me a review plan for what I missed over the next three weeks, with growing gaps." Put the dates in your calendar.
If you are studying for a course or certification we cover, the free Kestrel Exams practice apps are one more place to do retrieval practice.
Danger signs: when the AI is doing your learning#
Which session is more worrying? (a) It felt slow, you got several things wrong, and you needed hints. (b) Everything made sense as you read, and you finished in ten minutes.
Show the answer
(b). Smooth reading feels like learning but often is not, as the rereaders in the 2006 study showed. (a) is what effortful practice normally feels like.
| Sign | What it looks like | What to do instead |
|---|---|---|
| Copying answers | You paste the AI's output and could not redo it on a blank page. | Close the chat and redo it. Use the AI only to check. |
| Illusion of fluency | It reads so smoothly that you feel you know it. | Wait, then test yourself without looking. |
| Fixed, not checked | You asked "is this right?" and got a corrected version. | Repeat the rule: "Point to the first mistake. Don't fix it." |
| Trusting it blindly | You accept its "correct" without checking. | AI tutors make mistakes. Check key points by trying them or in a trusted source. |
The illusion of fluency is the "I get it" feeling that comes from reading something clear. It often tracks how easy the text was to read, not whether you can do it tomorrow. In the 2006 study, the rereaders were the most confident and, a week later, remembered the least. In the 2025 study, students with the plain assistant did not feel they had learned less, but they had. The honest test: close the chat and do one problem from a blank page.
For Priya, that means a new sheet with her own made-up orders and the formula written without the chat open. If she can do it and explain why each range goes where it does, the skill is hers.
Any subject, any job skill, and the rules you work under#
A nursing student, a Spanish learner and a new hire learning a work tool all want to learn with AI. What is the same about how each should set it up?
Show the answer
They produce something first. The AI checks, hints or asks instead of doing it. And they come back later to recall it, with growing gaps.
| Learner | What they produce first | What they ask the AI |
|---|---|---|
| Nursing student | A from-memory explanation of how the heart moves blood around the body | "Find the gaps. Don't rewrite it." |
| Language learner | Five sentences in the past tense | "Number my mistakes, but don't correct them yet." |
| New hire | The steps of a task, written from memory | "Quiz me on my training notes." (Leave out anything confidential.) |
| Beginner programmer | Their own attempt at a small program | "Give me one input where it fails. Don't show the fix." |
Priya's skill is a work skill, not a school subject, and nothing in the method changed: you produce, the AI checks.
Follow your instructor's rules, even when your AI use feels like harmless tutoring. Some classes allow AI for practice but not graded work; some ban it entirely; some ask you to say how you used it. If the rules are unclear, ask before you start. At work, check your employer's policy, especially before pasting company information into an AI tool. Module 10, AI at work, goes deeper on disclosing your AI use.
What was new in this module#
- Learning is being able to do it later without help, and it is built by retrieval: pulling ideas out of your own memory.
- Recalling beats rereading after a delay (the testing effect), especially with feedback, even though rereading feels better.
- Spaced practice beats cramming, and the best gap is longer the longer you need to remember.
- In one field experiment, a plain chat assistant raised practice scores but lowered exam scores; a hint-giving tutor avoided that harm.
- Seven tutor moves, from the hint ladder to a spaced review plan.
- Predict the answer first, and use a sanity check to spot impossible results.
- The danger signs, led by the illusion of fluency, and the blank-page test.
Make the AI hint, ask and check, so that you do the remembering. A finished answer you read is not a skill you have; the effort of recalling it is where the learning happens.
Key terms#
- Retrieval
- Pulling an idea or skill out of your own memory, without looking, and putting it to use.
- Testing effect
- The finding that trying to recall something strengthens the memory, not just measures it.
- Spacing (distributed practice)
- Spreading study sessions over days and weeks instead of cramming them together.
- Illusion of fluency
- The "I get it" feeling from reading something clear, which is not the same as being able to do it later.
- Meta-analysis
- A study that pools the results of many studies to find the average effect.
- Hint ladder
- Asking for the smallest useful hint first, and a bigger one only if you are still stuck.
- Socratic questions
- Questions, asked one at a time, that lead you to the answer.
- Explain-back
- You explain an idea in your own words; the AI finds gaps instead of rewriting it.
Check yourself#
Answer each one out loud before you open it. Reading the answer without trying first feels like learning and is not.
1. A coworker says: "I read the AI's explanation three times and now I totally get it." What would you suggest?
Close the chat, wait a while, and try a fresh problem from a blank page. Rereading builds confidence faster than memory: in the 2006 study the rereaders were the most confident and remembered the least a week later.
2. You ask your AI tutor to check your work, and it sends back a corrected version. What do you change?
Restate the rule: say whether it is right and point to the first mistake, without fixing it. Then fix it yourself. A correction is a finished answer, which skips the effort that makes a skill stick.
3. In the experiment published in 2025, practice scores with AI rose by 48% or more. Why is that not good news on its own?
Those scores were earned with the AI's help. What matters is what students could do without it, and the plain assistant group scored 17% lower on the unassisted exam.
4. You have an exam in six weeks and about three hours a week to study. How would you plan it?
Short sessions spread across the weeks, each starting with recall, with missed topics coming back sooner and known ones after longer gaps. Do not save it all for the last week.
5. Your class policy says "no AI on homework." Is the tutor prompt fine on homework, since it only gives hints?
No. The tutor setup does not change the rule. If you want AI practice on ungraded problems, ask your instructor first.
I can#
Tick each one only when you could do it tomorrow without this page open.
- I can explain why reading a finished answer is not the same as learning.
- I can set up any AI assistant as a tutor that hints, asks and checks instead of answering.
- I can use explain-back and quizzes on my own notes, and check the AI's feedback.
- I can plan spaced reviews and put them in my own calendar.
- I can spot the illusion of fluency and run a blank-page test.
Frequently asked questions#
Is it cheating to use AI to study?
It depends on your class or workplace rules, so check them first. Using AI for hints or quizzes is often allowed for practice, but some classes ban AI entirely. If the rules are unclear, ask before you start.
Does using ChatGPT or other AI make you learn less?
It can if it does the work for you. In one large experiment, students who practiced math with a plain chat assistant scored 17% lower on a later test without it. Students whose AI gave hints instead did about as well as students with no AI.
What is a good prompt to make AI act as a tutor?
Tell it not to give the final answer unless you ask. Ask for the smallest hint first and one question at a time. Have it check your attempts without fixing them, and have you explain ideas back in your own words.
What is retrieval practice?
Studying by recalling from memory, through quizzes, flashcards, practice problems or explaining without notes. Research shows it beats rereading for long-term memory.
How do I make an AI quiz me on my notes?
Paste notes you wrote, with nothing personal or confidential. Ask it to quiz you one question at a time, using only your notes, to explain your mistakes, and to list what you missed at the end.
This course is free, with no ads and no account. If it helped you, you can help keep it that way.
Support Kestrel Exams →Sources#
Roediger, H. L., III and Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249-255. doi.org/10.1111/j.1467-9280.2006.01693.x. Recall versus rereading results and the confidence finding.
Rowland, C. A. (2014). The effect of testing versus restudy on retention: A meta-analytic review of the testing effect. Psychological Bulletin, 140(6), 1432-1463. doi.org/10.1037/a0037559. 159 effect sizes from 61 studies; feedback and recall moderators.
Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T. and Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354-380. doi.org/10.1037/0033-2909.132.3.354. Spacing review.
Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J. and Willingham, D. T. (2013). Improving students' learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4-58. doi.org/10.1177/1529100612453266. Ratings for ten techniques.
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi.org/10.1073/pnas.2422633122. Classroom-randomized field experiment (an earlier working-paper version was titled "Generative AI Can Harm Learning").