Skip to content

AI Is a Better Tutor Than You Can Afford — If You Stop Asking It for Answers

The same model makes you learn twice as fast or half as much, and the difference is entirely method: as an answer machine it manufactures fluent illusions of understanding; used to increase your struggle and expose your confusion, it becomes the tutor almost no one could buy.

By Mehdi9 min read
Share
On this page

The same AI model will make you learn roughly twice as fast or half as much, and nothing about the model decides which — your method does. Used one way, it is an answer machine that hands you clean explanations you nod along to and forget by Thursday, manufacturing a fluent illusion of understanding that collapses the moment you have to produce anything yourself. Used the other way — against its own grain, to increase your struggle rather than remove it, to expose your confusion rather than paper over it — it becomes the thing Benjamin Bloom showed was worth two standard deviations of improvement and that almost no human could afford: a patient, personal tutor available at 2 a.m. for the price of a subscription.

This is a guide to the second way. It is not motivational. Each practice below is tied to a specific finding about how memory and understanding are actually built, because the practices only make sense once you see that the AI's default behavior is optimized for the wrong outcome, and you have to deliberately fight it.

The trap: fluency is not understanding, and the model sells fluency

Start with the mechanism the whole essay hangs on. There is a large, replicated body of work in learning science — much of it associated with Robert Bjork — on what he named desirable difficulties: conditions that slow you down and make acquisition feel harder in the moment, yet produce dramatically better long-term retention and transfer. Retrieval practice, spacing, interleaving, generating an answer before you're told it. The unifying and deeply counterintuitive result is that the feeling of learning and the fact of learning are often inversely related. A concept you read once in a clean explanation, understanding every line, feels learned. Two weeks later you cannot reconstruct it. The ease was the problem.

Now notice what an AI does by default. You ask, it explains — fluently, patiently, at exactly the level you request, with no friction at all. That is the single most seductive and most useless mode for actual learning, because it maximizes momentary fluency, the one metric that predicts retention negatively. You finish the conversation feeling you understand quantum tunneling or async runtimes or the bias-variance tradeoff. You have, in fact, watched someone else understand it. The illusion is total precisely because the explanation was good.

So the governing principle is uncomfortable: to learn well with AI, you must use it to make the experience harder, not easier — to add back the difficulty its default behavior strips away. Everything below is a way to do that.

1. Make it Socratic: turn the oracle into a questioner

The highest-leverage change you can make is a standing instruction that inverts the tool's default: ask me questions, one at a time, and do not give me the answer until I've genuinely tried. Turn the answer machine into a questioner.

The mechanism is the generation effect and the testing effect, two of the most robust results in the field (the testing-effect work is associated with Roediger and Karpicke). Retrieving or producing an answer yourself — even a wrong one, even a partial one — builds far stronger memory than reading the correct answer. A question forces retrieval. An answer permits recognition, which feels like knowing and isn't. When the model asks "before I explain, what do you think happens to the electron's probability amplitude at the barrier?" and waits, it has converted a passive read into an active retrieval, which is the whole game.

Concretely, I keep a saved prompt along the lines of: Act as a Socratic tutor. Ask me one question at a time to build up the concept. When I answer, tell me what's right and wrong in my reasoning, then ask the next question. Never explain something you could ask me to work out. The difference in what sticks, versus asking the same model to "explain X," is not subtle.

2. Generate before you consume

Before you let the AI explain anything, attempt it yourself — the problem, the derivation, the guess at why the code broke — and only then have the model check and correct you. This is the same generation effect applied at the level of workflow, and it is the practice people skip most, because attempting first feels inefficient when a correct answer is one keystroke away.

The efficiency intuition is exactly backwards. The value isn't in producing the right answer; it's in the act of reaching for it, which is what encodes memory and, crucially, what reveals the shape of your wrong model. When you guess that the API returns a list and it returns a generator, the surprise at being corrected is a far stronger encoding event than reading "it returns a generator" cold. You have to have committed to a prediction for the correction to land. Write your answer down — actually down, not in your head — then paste it in and ask the model to find the flaws.

3. Use it to find your confusion — the thing a textbook can't do

Here is the use that justifies the whole tool, and the one almost nobody reaches for: diagnosing precisely what you don't understand and why. Not "explain topic X," but "here is my current mental model of X, stated in full — where exactly is it wrong, and what's the smallest example that would break it?"

A textbook cannot do this. A textbook is a fixed artifact that explains the concept as it is structured, with no access to the specific misconception sitting in your head. A confusion is personal: it lives at the exact seam where your existing model meets the new material, and no author can anticipate which seam is yours. This is a large part of what made one-to-one tutoring worth two sigma in Bloom's work — a human tutor watches you, notices the precise place your reasoning wobbles, and aims there. The model can do the same if you feed it your reasoning instead of just your question. Say what you think is true, in detail, and ask it to locate the load-bearing error. The best twenty minutes you can spend on any hard topic is getting the machine to tell you, specifically, what you're confused about — because you usually don't know.

4. Self-test relentlessly

Have the AI generate practice problems and quiz you, then struggle with each one before you look at anything. Ask for problems slightly beyond what you've shown you can do, ask it to withhold solutions until you've committed an attempt, and ask it to vary the surface features so you can't pattern-match.

This is retrieval practice, the operational core of desirable difficulties, and its power is that testing is not measurement of learning — it is learning. Every time you successfully pull a concept out of memory against resistance, you make the next retrieval easier and more durable; the struggle to retrieve is the strengthening event. The AI removes the historical bottleneck, which was that generating good, varied practice problems and grading your attempts took a teacher's time. A model will produce fifty variations and diagnose each of your answers for free. Interleaving — mixing problem types so you must first figure out which method applies, the hardest and most transferable skill — is a one-line request most people never make.

5. Go to the primary source for anything you actually need to know

Use the AI to orient, map the terrain, and find what matters — then read the original for anything you need to truly own. The model is a superb index and a mediocre authority, and the two roles must not be confused.

The reason is structural, not a matter of model quality, and I've argued it at length in Reading Primary Sources Is Becoming a Superpower: a summary is lossy compression optimized to preserve the gist, which means it systematically deletes the caveats, effect sizes, and boundary conditions that are the entire content of real understanding. An AI explanation is a compression of those compressions. For the ninety percent of things you're merely orienting yourself around, that's fine and efficient. For the ten percent you will build on, decide with, or stake your name on, the stripped caveats are exactly what you needed, and only the source has them. Let the model tell you which three papers matter and why. Then read the three papers.

6. Verify what it teaches, because it can be confidently wrong

AI models produce false statements in the same fluent, assured register as true ones — there is no tremor in the voice when the model invents a citation or misstates a theorem. So any claim you're going to rely on gets cross-checked against a source that isn't the model.

The reason this can't be prompted away is that fluency is not correlated with correctness; confident phrasing is the default output style regardless of whether the underlying claim is grounded. This does not make the tool untrustworthy for learning — it makes passive consumption of novel facts the one mode to avoid, which is why the other six practices keep you active. When you're generating and it's checking you, or asking and you're retrieving, an error is cheap and often self-revealing. When you're banking its assertions as fact, an error propagates silently into everything you build next. Verify the load-bearing claims. Let the trivia go.

7. Explain it back — the Feynman move, with an audience that talks back

Finally, the test that catches every illusion the other practices might miss: teach the concept, in plain language, to the AI. If you can't — if you reach for a word you can't unpack, or a step you can only gesture at — you have found the edge of your actual understanding, which is never where you thought it was.

This is the technique associated with Richard Feynman, and it works because comprehension and production are different capacities that feel identical from the inside. You can follow an argument perfectly and be unable to generate it, and the gap between those two is precisely where borrowed understanding hides. Explaining forces production. What the AI adds is the part that used to be scarce: an inexhaustible audience that interrogates you. Tell it to play a bright, confused student and ask the follow-up questions your explanation invites. When it asks "wait, why does that step follow?" and you can't answer, the illusion breaks in the one way that's useful — at the broken step, by name.

The honest counterargument

The strongest case against all of this is efficiency, and it deserves a straight answer. If the goal is to get the task done — ship the code, write the memo, produce the analysis — then struggling first is pure overhead, and the rational move is to let the model do it and move on. That's often correct. Not everything is worth learning; a great deal is worth merely outsourcing, and pretending otherwise is its own kind of waste. The friction I'm prescribing is justified only for the things you've decided to carry in your own head — because you'll compose them, judge them, or be the person in the room when the model isn't. Which things those are is the genuinely hard question, and it is a different essay: What to Learn When AI Can Explain Anything. This guide assumes you've already decided something is worth learning. It only claims that once you have, the answer-machine reflex will quietly sabotage you.

There's a cleaner way to hold the whole thing. When Kasparov lost to Deep Blue in 1997 and then built "advanced chess," the lesson people took was that human-plus-machine beats machine alone. The subtler lesson was that the pairing wins only when the human brings something the machine doesn't — judgment the human actually holds, not judgment the human has outsourced and is pretending to have. An AI you use to avoid thinking makes you a worse centaur, not a better one. An AI you use to think harder than you otherwise would makes you the tutored student in Bloom's experiment, at a price the experiment couldn't imagine.

The machine will always offer you the answer. Learning is the discipline of asking it for the question instead.

Frequently asked questions

Doesn't making learning deliberately harder just waste time when the AI could give me the answer in seconds?
It feels like waste and isn't, and the confusion is exactly what Bjork's term 'desirable difficulties' names. Fast, fluent acquisition and durable retention are close to opposites: the conditions that make a concept easy to absorb in the moment — a clean explanation you read once, nodding — tend to produce the weakest long-term memory and the worst transfer to new problems, while the conditions that feel effortful (retrieving before you look, generating before you're told, struggling with a problem you can't yet solve) build memory that lasts. The AI-given answer optimizes the metric you can feel — momentary fluency — which is the wrong one. The extra minute you spend attempting before checking is not tax; it is the mechanism. If you only have thirty minutes, spend more of them retrieving and less of them reading.
If the AI is confidently wrong sometimes, how can I trust anything it teaches me?
You don't trust it — you use it in the two modes where being wrong costs least and helps most. As a questioner it produces prompts, not facts: a bad question wastes ten seconds and a good one exposes a real gap, and neither depends on the model being correct. As a checker of work you already generated, you're comparing its account against your own reasoning, which surfaces disagreements you can then resolve at the source. The danger zone is the third mode — passive consumption of novel factual claims you can't evaluate — and the rule there is simple: for anything load-bearing, cross-check against a primary source. Treat the model as a fast, fallible tutor whose errors you catch by staying active, not an oracle whose output you bank.
Isn't the Feynman technique just a study tip? Why does explaining it back matter so much with AI specifically?
Explaining a concept in plain language is the cheapest available test of whether your understanding is real or borrowed, because comprehension is deceptive and production is not — you can follow an argument line by line and still be unable to reconstruct it, and the gap between those two states is precisely where the illusion of understanding lives. What AI adds is an inexhaustible, non-judgmental audience that talks back: you teach the concept to the model, and it can play a confused student, ask the follow-up that finds your hand-waving, or push on the step you glossed. A textbook can't notice that your explanation broke down at step three. A patient human tutor could, but almost no one has one on demand. That specific interaction — being interrogated on your own explanation until the gaps show — is the thing that used to be scarce and now isn't.

Filed under Future & Modern Skills. The capabilities that stay valuable as the tools change.

Essays like this, in your inbox.

Thoughtful essays. No spam. Unsubscribe anytime.

Future & Modern Skills

Productive Struggle Is What AI Removes

Understanding is built by effortful retrieval and self-generated answers, not by receiving them. An AI that hands you the answer removes exactly the struggle that creates the learning — so the more helpful it feels, the less you keep.

8 min read
Future & Modern Skills

When AI Can Explain Anything, Learn the Things It Can't Hand You

Free explanation doesn't make learning worthless — it moves the value from acquiring facts to what a model can't give you: judgment, tacit skill, and enough real expertise to tell a correct answer from a confident wrong one.

8 min read
Future & Modern Skills

The Centaur and the Cyborg: The Meta-Skill Is Knowing Which Mode a Task Needs

There are two distinct ways to work with an AI — the centaur, who keeps a clean seam and delegates whole sub-tasks, and the cyborg, who dissolves the seam and thinks in a tight loop. Choosing correctly per task, and building the verification each mode demands, is the core professional skill.

9 min read