How to Use an AI Tutor Without Letting It Do the Thinking

I teach adult students in a Montreal classroom by day and build an AI tutor the rest of the time. That gives me a front-row seat to a quiet problem: students who use AI every day feel like they are learning more than ever — and many are learning less than ever. The tool is not the problem. The way most people use it is. What follows is the protocol I give my own students. It works with any AI, including mine.

The Answer Trap

Ask an AI to solve your problem and it will, instantly, in clean confident prose. You read the solution, it makes sense, you move on. It feels like learning. It is the opposite of learning. What you actually practised was reading. What the exam will demand is producing — from a blank page, under time pressure, with no chat window in sight.

Cognitive science has been unusually clear about this for decades. Memory research shows that retrieving an answer yourself strengthens long-term retention far more than reading that same answer — a finding known as the testing effect [1]. A major review of study techniques reached the same verdict: practice testing sits at the top of the list of what works, while re-reading — which is what consuming AI answers amounts to — sits near the bottom [2].

Why Struggle Is the Point

Robert Bjork calls the underlying principle desirable difficulties: conditions that make practice feel slower and harder often produce stronger learning, and conditions that make practice feel smooth and easy often produce weaker learning [3]. The struggle to pull an idea out of your own head is not an obstacle on the way to learning. It is the mechanism itself.

This is exactly what a general-purpose chatbot removes. It is optimized to satisfy you in the moment — fast, fluent, frictionless. Nobody tuned it for what you can recall three weeks later, alone, in an exam room. The interests of the tool and the interests of your memory are not aligned by default. You have to align them yourself.

The Protocol: Six Moves

So here is the protocol. One: attempt first. Before you ask the AI anything, write your own answer — even a bad one, even two lines. The attempt commits you to a position, exposes what you actually do not know, and primes you to notice exactly where the explanation diverges from your thinking. An attempt followed by feedback is worth far more than feedback alone.

Two: ask for questions, not answers. Tell the AI to quiz you. Ask it “what should I be asking myself here?” instead of “what is the answer?” Three: explain it back. Teach the concept to the AI in your own words and ask it to find the holes in your explanation. Self-explanation is one of the best-documented ways to turn exposure into understanding [4] — and an AI is a tireless, judgment-free audience.

Four: chase the why. Follow every answer with why, and then why again, until you can reconstruct the whole chain without the machine. If a step only holds because “the AI said so,” you do not have it yet. Five: let errors surface. Do not sanitize your attempts before showing them. A wrong attempt that gets corrected teaches you more than a right answer that gets copied — the error marks the exact spot where your model of the subject was broken.

Six: space it. Come back tomorrow, without the chat history, and redo the problem cold. Spaced retrieval is one of the most robust effects in the learning literature [5]. If you cannot do it a day later on a blank page, you never had it — you were renting the understanding from the machine.

Red Flags You Have Gone Passive

How do you know you have slipped into passive mode? Three red flags. You cannot redo the problem alone. The solution made perfect sense while you read it, and none of it is available when the window closes. You read fluently but recall nothing. Fluency is the great impostor of studying; it feels like knowledge and measures only familiarity. Your prompts start with “do” instead of “check.” “Do this problem” outsources the thinking. “Check my reasoning” keeps it where it belongs.

The self-test is brutal and takes five minutes: close the chat, take a blank page, and redo today’s hardest problem. Whatever survives the blank page is yours. Whatever does not was never learning — it was watching.

The Hallucination Problem

There is a second reason blind copying is dangerous, and it has nothing to do with pedagogy. Large language models generate text that is statistically plausible, not verified — and they state their fabrications with the same fluent confidence as their facts [6]. A student who copies without thinking absorbs the errors along with the answers, and cannot tell which is which. Verify against your course materials. Your textbook and your teacher’s notes define what your exam considers true; the chatbot does not.

The protocol protects you here too. If you attempted the problem first and chased the why, a hallucinated step does not slide past you — it collides with your own reasoning and with your notes, and the collision is exactly the signal to slow down and check. Passive readers absorb errors. Active learners catch them.

What a Good Tutor Refuses to Do

None of this is hypothetical to me. When we built our tutor, we made a decision that costs us something with every impatient user: it does not hand over answers. It asks what you have tried, guides you a step at a time, and holds back the solution while you can still reach it yourself. Some students hate this for the first week. It is the same discipline as this protocol, enforced by the software instead of by willpower.

But most students, most of the time, will be working with a general-purpose chatbot that will happily do their thinking for them at the first request. Those tools will not protect your learning, because they were never designed to. The discipline has to live in you. The six moves above are how you carry it into any chat window, on any platform, with any model.

The AI is not the threat. Outsourced thinking is. Used as an answer machine, the most powerful learning technology ever built will quietly make you weaker. Used as a sparring partner — one that questions you, quizzes you, and checks your reasoning — it is the best study partner you have ever had. The difference is not in the model. It is in the first sentence you type.