How to Use ChatGPT to Study (So It Actually Sticks)

Here is how most people study with ChatGPT. They paste in a paragraph they don't get, ask it to explain, read the explanation, think "oh, that makes sense," and move on.

It does make sense. The explanation is clear, patient and tailored to the exact thing you were confused about. Then, a week later, a practice question comes at the same idea from a slightly different direction and you have nothing.

The explanation wasn't the problem. The problem is that understanding something when it's explained to you and being able to produce it on your own are two different skills, and an exam only tests the second one.

So the single most useful change you can make is this: stop asking ChatGPT to explain. Ask it to quiz you.

Why being asked beats being told

Cognitive scientists call it retrieval practice, and it is one of the most replicated findings in learning research. Pulling an answer out of your own memory strengthens that memory far more than reading the answer again.

In a well-known 2006 study, Roediger and Karpicke had students either read a short text four times, or read it once and then try to write down everything they remembered three times. Tested five minutes later, the rereaders did slightly better. Tested a week later, the recall group remembered 61% and the rereaders 40%. The rereaders, asked beforehand, had been the most confident.

An AI explanation is the most pleasant kind of rereading ever invented. That's exactly why it can fool you.

Five prompts to copy

Each of these works in ChatGPT, Claude, Gemini or any similar assistant. Replace the bracketed parts.

1. Quiz me, one question at a time

I'm studying for [exam / course]. Below is a section of my material.
Don't explain it yet. Ask me 5 questions about it, one at a time, and wait for my answer before moving on.
After each answer, tell me if I was right and quote the sentence from the text that settles it.
If I'm right, don't praise me — just ask the next question.
At least 2 of the 5 should use a scenario that doesn't appear in the text.

[paste the section]

"One at a time" stops it from dumping five questions and their answers in a single message. "Quote the sentence" keeps it anchored to your material. "Don't praise me" removes the warm feeling that tells you you're done when you aren't.

2. Same idea, new angle

Getting a question right once may only mean you remember that question.

I just got this question right:
[question and my answer]

Write a new question that tests the same idea from a different angle — different numbers, different situation, or ask what happens if one condition changes.
Give me only the question. Wait for my answer before you show yours.

If you get the new one wrong, you didn't know the idea. You knew the question.

3. Grade my explanation, don't rewrite it

Here is my explanation of [concept], in my own words.
Point out: (1) anything that's wrong, (2) anything important I left out, (3) any sentence where I seem to be repeating the textbook rather than showing I understand it.
Do not rewrite it for me. I'll revise and send it back.

If it rewrites your answer, you'll read the polished version, nod, and learn nothing. Revising it yourself is the part that does the work.

4. Tell apart the two things you always mix up

I keep confusing [A] and [B].
Write 4 short scenarios that differ by one key detail, and for each ask me whether it's A or B.
One at a time. After each, tell me which detail decided it.

5. Mix the chapters

Here are my notes from chapters [X], [Y] and [Z].
Quiz me with 8 questions drawn from all three, in random order, without telling me which chapter each one comes from.

On the exam nobody tells you which chapter a question belongs to. Working out what kind of problem this is is half the skill, and practising one chapter at a time hides it from you.

Three traps

It makes things up, confidently. Specific numbers, citations, section numbers and page references are where AI is most likely to be wrong while sounding certain. Give it your own material, tell it to answer only from that and say so when the text doesn't cover something, and check anything that matters against the source.

It agrees with you. Ask "it's B, right?" and you'll often get "Yes, B is correct" even when it isn't. Give answers without leaning, and add a line like "Give your verdict first, and don't change it because of how I phrase things."

It doesn't come back. This is the one people miss. The question you got wrong on Tuesday is exactly the one you need to be asked again — the next day, then a week later, then a couple of weeks after that. A chat won't do that on its own. Start a new conversation and it has no idea what you missed.

A simple routine

1. Paste in one section. Run prompt 1.
2. For anything you miss, ask for a short explanation of just that point, then run prompt 2.
3. Write every miss in a spreadsheet with today's date.
4. Each day, before starting anything new, re-answer the misses that are due: next day, then about a week, then about two weeks, then a month. Get one wrong and it goes back to the next day.

That's it. ChatGPT handles the explaining and the question-writing; you do the answering; the spreadsheet does the remembering.

If keeping that spreadsheet is the part you know you'll drop, that's the gap Tali was built for: it writes the questions from the file you upload, keeps track of what you miss, and brings it back on a schedule. But the routine above works with nothing more than a chat window and a sheet.

The test of a good study session

At the end of a session with ChatGPT, ask yourself one question: who did most of the talking?

If it was the AI, you had a very good lecture. If it was you, you studied.