Discover strategies on how to use AI as a tutor, focusing on techniques that promote active learning and improve your grades.
Short answer: make the AI ask the questions. If the chatbot is doing all the answering, you aren’t studying. You’re watching.
That’s most of the article. The rest is why, plus the exact prompts I’d use if I had to cram for an organic chemistry final again. (I got a C+. Take the advice anyway.)
The Study That Should Scare You a Little
In 2024, Wharton researchers ran an experiment with nearly a thousand Turkish high school students doing math. One group got plain GPT-4 during practice sessions. Their practice scores jumped 48 percent. Great, right?
Then came the exam, with no AI allowed. That same group scored 17 percent worse than kids who never had AI at all.
Read that again. The tool made them feel smarter while making them worse at math. It’s the cleanest example of AI hype vs reality I’ve seen, because nobody lied to those students. They used the thing the obvious way: paste the problem, copy the answer, feel productive.
A second group got a tweaked version the researchers called GPT Tutor. It gave hints and wouldn’t hand over full solutions. Their practice scores rose 127 percent, and their exam scores didn’t drop.
Same model, different rules, and a completely different result.
So when someone asks whether AI is worth it for studying, the real answer is annoying. It depends on who runs the conversation. If the chatbot runs it, you lose. If you run it, and you’ve told it to hold back, you get the most patient tutor who ever lived, one that’ll explain the Krebs cycle a ninth time at 1 a.m. without sighing.
Set the Rules Before You Ask Anything
Default chatbots are people-pleasers. Ask one a question and it sprints to the answer, bullet points and all. You have to break that habit in your first message.
Here’s a setup prompt I’d actually paste:
“You’re my tutor for [subject]. Don’t give me answers until I’ve tried twice. Ask one question at a time. If I’m wrong, give a hint, not the fix. Keep replies under 100 words.”
That last line matters more than it looks. Long replies feel generous, but you skim them. And skimming is just reading with extra guilt.
Most of the everyday AI tools you already have, like ChatGPT, Claude and Gemini, will follow instructions like these if you’re explicit. Some let you save them as custom instructions or a project, so you don’t retype them every session. Do that. At midnight you won’t remember to be disciplined.
A quick tangent: tell it your level. “I’m a second-year nursing student” gets you very different explanations than “explain pharmacokinetics.” Without that context, it pitches everything at some imaginary average person. That person doesn’t exist, and they definitely aren’t sitting your exam on Thursday.
Make It Test You Until It Hurts
Here’s the part productivity YouTubers skip. Rereading your notes barely works. Testing yourself does.
This isn’t just a feeling. In 2006, psychologists Henry Roediger and Jeffrey Karpicke had students either reread a passage or get tested on it. A week later, the tested group remembered far more, even though they’d seen the material less. Retrieval is the workout. Rereading is watching someone else lift.
AI turns out to be weirdly good at this. Paste in your lecture notes, say 2,000 words on the French Revolution, and ask for ten short-answer questions. Skip multiple choice. It lets you recognize the answer instead of producing it, which is a polite way of cheating yourself.
Then answer in your own words, badly, and let it grade you.
A few prompts worth stealing:
- “Quiz me on this chapter. Start easy, and get harder after I get two right in a row.”
- “Give me a problem like this one, with different numbers.”
- “Make me explain this concept to a 12-year-old. Tell me where I got vague.”
That last one stings. You’ll find you can define “opportunity cost” but stall when asked for one example. Good. That gap is what you came to find.
Two days later, ask it to quiz you on the same material again. Spaced review beats cramming, and the chatbot doesn’t care that you asked before. It has no memory of your embarrassment. Lucky thing.
Catch It When It Gets Things Wrong
Chatbots get things wrong with total confidence. Not constantly, but often enough to matter.
I’ve watched one invent a page number in a textbook it had never seen. Math is another weak spot. In multi-step arithmetic a sign flips somewhere, and the derivative ends up wrong while the explanation still sounds lovely.
So treat it like a smart classmate who didn’t do the reading. Useful, but occasionally full of it.
Some practical rules:
- Check any fact that’ll show up on an exam against your textbook or slides.
- Your professor’s definition wins, even if the AI’s sounds better.
- For math, ask it to verify the answer by plugging it back in.
- If a citation looks convenient, search for it. Fake references are a known problem, and “the robot said so” won’t impress your TA.
One trick really does help: give it your own course material first. When answers come from notes you uploaded, it’s much less likely to make things up.
When Not to Use AI at All
Some tasks need your brain struggling on its own, and handing them off defeats the point.
First drafts of essays are the big one. The messy, crossed-out thinking you do while writing is where the learning happens. If a chatbot writes your first draft, you skip the part where you figure out what you think. Ask it to critique your draft instead, and tell it to be harsh.
Graded problem sets are another. Beyond the obvious academic-integrity problem (plenty of universities now check), you’d be repeating the Wharton students’ mistake on purpose.
Your first attempt at any hard problem should also be yours alone. Give it ten real minutes before you ask for a hint. Frustration is unpleasant, sure. It’s also when your brain does its most useful work of the whole session.
One less obvious rule: don’t let it decide what you study. Your syllabus, past papers and whatever the lecturer repeated three times have already told you. The AI doesn’t know your professor is obsessed with the Treaty of Westphalia. You do.
You Probably Don’t Need to Pay
Now the money question. For most students, the paid tier isn’t worth it.
You can get a long way using AI without paying for this kind of work. Quizzing, explaining and checking a proof are all AI for small tasks, and free tiers handle them fine. You’ll hit message limits during a marathon session. That’s annoying, not fatal. Switch to another free tool, or take it as a sign to go for a walk.
Paying helps in a few cases. One is uploading huge files, like a 400-page textbook PDF. Another is heavy use of the stronger reasoning models for advanced physics or math. If you’re in a quantitative degree and use it daily, maybe pay. First check whether your university already gives you access. Many do now, and you’re paying for it through tuition anyway.
Don’t pay because a TikTok promised the premium model will “10x your grades.” It won’t. You will.
Rehearse the Exam Without the Crutch
A day or two before the real thing, ask the chatbot to write you a mock exam. Give it past papers if you have them: “Three essay questions, 45 minutes each, in this style.” Then close the tab.
Sit it properly. Kitchen table, phone in another room, timer running, a cheap pen that skips. Only after that do you paste your answers back in and ask it to mark them against the rubric.
It’s the one part of the process where the AI should be completely absent, and it’s the part that shows you where you really stand.
