AI tutors and chatbots: use them without outsourcing your thinking#

Imagine the most patient tutor in the world. It never sighs when you ask the same question a third time. It will re-explain photosynthesis as if you were five, then as if you were a biochemist, then with a cooking analogy, all in the time it takes to type. It will invent a hundred practice questions on the spot and mark every one. That tutor now exists, it’s free or nearly so, and it’s sitting in a browser tab right next to the one you’re studying in.

Here’s the catch, and it’s a big one. The exact same tool, asked a slightly different way, becomes a machine for not learning—a way to make the work disappear so smoothly that you walk away feeling clever and remembering nothing. A large language model (LLM) will happily hand you the finished essay, the worked solution, the summary you were supposed to write yourself. And the moment it does the effortful part for you, it has quietly stolen the part where the learning actually happens. The skill is not in using AI. The skill is in deciding what to let it do.

The offloading trap#

There’s a well-studied habit behind this danger, and psychologists call it cognitive offloading: using something in the world—a notepad, a calculator, a phone, now a chatbot—to do mental work you’d otherwise do in your head. Risko and Gilbert’s review of the research lays out the trade-off cleanly. Offloading is often the smart move: nobody memorises a 30-digit number when paper is to hand, and nobody should. But offloading has a cost that shows up later. When you let an external tool hold the information or do the reasoning, you tend not to encode it yourself—so when the tool isn’t there, neither is the knowledge.

For a person trying to learn, that cost is the whole problem. If your goal were just to produce an answer, offloading to AI would be pure win. But your goal is to change what’s in your own head—to be able to do this yourself, next week, in the exam, in the meeting, when the tab is closed. Every time you ask the AI to do the thinking you came here to learn, you trade a durable change in you for a disposable result on the screen.

I want to be fair to the tool here. The danger isn’t AI; it’s a much older instinct. Left to ourselves we all drift toward whatever feels like progress with the least strain, and an LLM is the most frictionless source of that feeling ever built. Watching a perfect explanation scroll past feels like understanding. It usually isn’t—it’s recognition, the same fluency illusion that makes re-reading feel so productive and do so little.

Make it generate difficulty, not remove it#

The way out is to remember what makes learning last. Durable learning comes from desirable difficulties—effortful retrieval, spacing, wrestling something into your own words—not from smooth, easy intake. So point the AI at creating that productive struggle for you, never at sparing you from it. A good rule of thumb: the AI should be the thing that makes you work, not the thing that does the work.

Things to ask an AI tutor to do—these add effort:

Quiz me, don’t tell me. “Ask me ten questions on this chapter, one at a time. Wait for my answer before you reveal anything, then tell me what I got wrong and why.” Now the AI is running retrieval practice on you—the single highest-leverage study move there is—instead of handing you the answers to read.
Explain at my level, then make me explain it back. Get the explanation pitched right for you—“explain it like I know basic algebra but not calculus.” Then close it and try to teach the idea back to the AI in your own words. Ask it to catch what you fudged. The teaching-back is where the learning is.
Generate practice and feedback. Endless fresh problems, worked after you attempt them, with feedback on your specific mistakes. Attempt first, every time. Looking at the worked solution before you’ve struggled is the offloading trap wearing a study hat.
Find my gaps. “Here’s my explanation of how a recession spreads. What did I get wrong, what’s missing, and what would a sharp examiner poke at?” Use it as a mirror for the holes you can’t see, then go back and fill them yourself.
Be a Socratic sparring partner. Tell it: “Don’t give me answers. Ask me questions that lead me to work it out.” A surprisingly capable tutoring style, and one that keeps the cognitive load where it belongs—on you.

And the one thing to almost never ask: “Just write it / solve it / summarise it for me,” when it is the thing you’re trying to learn to do. Getting the answer is not the same as being able to produce the answer, and only the second one is learning.

Trust nothing, verify everything#

There’s a second reason not to let an LLM do your thinking, and it’s about the tool’s reliability, not just your effort. These models generate fluent, confident text—and they are perfectly capable of producing fluent, confident text that is wrong. The model can invent a citation, misstate a date, garble a formula, or fabricate a quotation, and it will do so in exactly the same authoritative tone it uses when it’s right. The industry word for this is “hallucination,” and there is no version of these tools, however advanced, that you should treat as an oracle.

For a learner this cuts two ways. The obvious risk is absorbing something false. The subtler risk is that you can’t catch a falsehood in a subject you don’t yet know—which is precisely the subject you’re using the AI for. So build a habit:

Treat AI output as a knowledgeable friend’s first guess, not as a source. It’s a brilliant starting point and a terrible final word.
Verify anything load-bearing against a real source—your textbook, course materials, a primary reference—before you trust it, and always before you cite it. Never repeat an AI’s citation you haven’t seen with your own eyes; it may not exist.
Use the friction as learning. Cross-checking the AI against your textbook is itself retrieval and elaboration—you’re not just fact-checking, you’re studying. The verification step isn’t overhead on top of the learning; some of it is the learning.

The good news is that the same scepticism that protects you from errors also protects you from the offloading trap. A learner who checks the AI’s work is, by definition, doing the thinking.

Takeaway#

An AI tutor is a power tool, and like any power tool it cuts in whatever direction you point it. Point it at removing effort—“do this for me”—and it will erode the very struggle that builds knowledge, with the added hazard of confidently feeding you things that aren’t true. Point it at adding effort—quiz me, make me explain it back, give me problems to attempt, show me my gaps—and it becomes the patient, infinite, personalised tutor that genuinely helps you learn. Ask yourself one question before every prompt: am I making the machine think, or making it think for me? Only the first one teaches you anything.