# Why Learn Anything?

A machine can already answer most of what is on this page. An experiment you run on yourself, two studies that disagree, and why I am building a tutor anyway.

By İbrahim Çoban, 3 November 2026. https://ibracob.dev/why-learn-anything

[Illustration: A small figure at a drafting table on a mountain ledge, drawing a map that unrolls into real mountains]

Before I argue anything, do this. It takes about forty seconds, and I will explain it later.

## The obvious question

If a machine keeps getting better at everything I could learn, why spend four years learning it?

It is a good question, and most of the answers are bad. "You need critical thinking" is a slogan. "AI makes mistakes" has an expiry date. I use AI coding tools to build my own software, so I am in no position to pretend the machine is not good.

So let's grant everything. Imagine an oracle. It is free, instant, and never wrong. You can ask it anything. Why learn?

## The oracle

Here is a map of everything you could know. You know one small part of it. You can only ask about what borders that part, because you cannot ask about a thing you have never heard of, in words you do not have.

Ask the oracle about everything at your edge. Watch the three numbers. Then switch to learning and press one cell three times.

The oracle gave you every answer you asked for, and you can ask exactly as many questions as you could before. Learning was slower and it moved the edge.

The strongest objection is that the oracle could also tell you what to ask, and good ones already do. I think that is true, and I think it changes what you are doing. You are no longer exploring. You are on a very good tour of a map somebody else is holding.

You can rent an answer. You cannot rent the next question.

That is an argument. Here is what happened when people tested something close to it.

## Two studies that disagree

In the fall of 2023, researchers ran a field experiment with nearly a thousand math students at a high school in Turkey. One group practiced with a plain chat assistant. Another used a tutor built to give hints rather than answers. A third had no AI at all.

During practice, the plain assistant group scored 48% better than students with no AI, and the hint tutor group 127% better. Look only at that and you would buy either product. Then take the AI away. On the exam the plain assistant group scored 17% worse than students who never had it, and the tutor group about the same. Their practice scores had been real, and they had not been learning.

> Note: One subject, in one school, in one country. It is strong evidence and it is not the last word.

Now the other study. In 2016 AlphaGo beat Lee Sedol, and the strongest Go player on Earth stopped being a person. If "why learn" ever had an obvious answer, it was there.

Shin and colleagues scored 5.8 million professional moves from 1950 to 2021 against a superhuman program. Human decisions got better after superhuman AI arrived, and players started making moves nobody had played before.

> Note: The authors are careful here. This is a pattern over time, not a controlled experiment.

So the same kind of machine made students worse at math and players better at Go. My reading of the difference is simple. At a Go tournament, nobody can hand their move to AlphaGo. They study what it did, and then they have to play.

The machine was not the variable. Who made the move was.

## So I am building a tutor

That is a strange conclusion to reach while building an AI tutor. It is called Folio, and here is why I think it can work. I will keep what I know apart from what I do not.

**You stay on your page.** Most students leave their PDF to ask an AI. Folio opens the lecture PDF, or a GoodNotes or Notability notebook with the handwriting still editable. You circle what you do not understand with the Apple Pencil.

**The tutor marks the page and talks.** Its voice and its drawing arrive together, on the passage they are about. You do not have to translate "the second term" back into a place on the page.

**Longer working goes on a facing sheet.** A matching number ties it to the passage it explains, and your original page stays readable. The work stays with the document, so you can come back to it.

**Then you can hide it.** Any explanation can be hidden and brought back. Hidden, the page is a test again, and the move is yours.

A sketch drawn for this essay, not a recording of the app.

Three reasons I expect this to work, in order of how much I trust them.

**The explanation is in the right place.** Richard Mayer's research on multimedia learning has found for decades that people learn better from words next to the picture they describe, and from narration heard while the picture is drawn. A chat window in another app breaks both of those. This research is about learning materials in general, not about Folio.

**The student is already working.** The question comes from a page they were studying, circled with their own hand. That matters, because the Turkish study suggests the danger is not the AI's answer. It is when the answer replaces the attempt.

**Friends are in the room.** A study room shares the document, the ink and the tutor, and keeps one set of notes from what the group says. Explaining a thing to a friend is the oldest way of making the move yourself, and in a room the tutor is one voice among several.

## What I do not know

I have shown the prototype to more than fifty students at my school, and roughly 80% asked where they could download it without my asking whether they wanted it. That is interest. My last product had fifty businesses on its waitlist, and it taught me not to mistake that for use.

I have not measured whether students come back. I have not measured whether they learn.

Here is how this fails. A tutor that explains beautifully can be the plain chat assistant with better drawings. Practice goes up, everyone is happy, and the exam goes down. Nothing I have built so far proves Folio avoids that. Hiding an explanation makes the attempt possible. It does not make anyone take it.

So the first test is whether students return within seven days. The test I actually care about is harder, and it is the one the Turkish researchers ran: what can they do with the app closed?

## Your eight words

Remember the pairs at the top. You read four of them and completed four. Type the second word of each, without scrolling up.

Completing a word took a few seconds more than reading one. In 1978 Slamecka and Graf showed that people remember words they produce better than words they only read, and a later review of 86 studies found the advantage is modest and reliable. Eight words in an essay prove nothing about you, in either direction. If your numbers went the other way, you are not an exception to anything.

But it is the whole argument at a size you can hold: the words you had to make are the ones most likely to stay.

The machine will keep improving, and I have no argument against that. But every answer it gives lands somewhere, and what it can do for you depends on the map it lands on. That map only gets drawn one way, and it takes three tries.

**Sources.** Bastani et al., "Generative AI without guardrails can harm learning," [PNAS, 2025](https://doi.org/10.1073/pnas.2422633122), with figures as reported by [Wharton](https://knowledge.wharton.upenn.edu/article/without-guardrails-generative-ai-can-harm-education). Shin et al., "Superhuman artificial intelligence can improve human decision-making by increasing novelty," [PNAS, 2023](https://arxiv.org/abs/2303.07462). Slamecka and Graf, "The generation effect," Journal of Experimental Psychology, 1978. Bertsch et al., "The generation effect: A meta-analytic review," [Memory & Cognition, 2007](https://mcdaniel97.github.io/Publications/Bertsch%20et%20al.%202007.pdf). Mayer, "Using multimedia for e-learning," Journal of Computer Assisted Learning, 2017.
