Teach it back: the fastest way to find out what you don’t know

Zips and toilets

How does a zip work? Most people, asked to rate their understanding on a scale of one to seven, give themselves a respectable score. They have used thousands of zips. They have watched the teeth come together.

Now explain it, step by step, in writing. What exactly does the slider do to the teeth? Why don’t they come apart when you pull on the fabric? Why does a zip sometimes split behind the slider?

That is roughly what Leonid Rozenblit and Frank Keil did in a series of studies published in 2002. People rated their understanding of everyday devices — zips, flush toilets, cylinder locks, speedometers — then wrote explanations, then answered diagnostic questions, then rated themselves again. The ratings dropped, often a long way. People were surprised by how little they could actually explain. The researchers called it the illusion of explanatory depth, and they found it was much stronger for explanations of how things work than for facts, procedures or stories, where people judged themselves more accurately.

The finding explains a familiar experience. Reading a clear explanation produces a feeling of understanding that is real as a feeling and unreliable as a measure. The only way to check it is to try producing the explanation yourself. And the most natural way to produce an explanation is to teach.

Expecting to teach

The benefit starts before you teach anyone. In a study by John Nestojko and colleagues, two groups read the same passage. One was told they would be tested on it; the other was told they would have to teach it to another student. In fact both were simply tested. The group expecting to teach recalled more, recalled the main points better, and organised what they recalled more coherently (Nestojko et al., 2014).

The likely reason is that teaching is a different goal from remembering. If you expect a test, you try to hold on to the material. If you expect to teach, you try to work out how it fits together, what comes first, and which parts matter — because you’ll need to explain them to someone who doesn’t already know.

Actually teaching

Expecting to teach helps. Doing it helps more, and for longer.

Logan Fiorella and Richard Mayer had students learn how the Doppler effect works. Some studied for a test. Some prepared to teach. Some prepared and then actually taught, by recording a short lecture. On an immediate test, preparing to teach gave an advantage. A week later, only the students who had actually taught were still ahead (Fiorella & Mayer, 2013; see also their 2014 follow-up).

A meta-analysis by Keiichi Kobayashi pulled the evidence together. Across the studies, preparing to teach produced a modest benefit, with an effect size of about g = 0.35. Teaching after preparing produced a larger one, about g = 0.56. And the effect was larger again when the teaching was interactive — when the learner had to answer questions from the person being taught, rather than just delivering an explanation into a camera (Kobayashi, 2019).

That last point is the key to doing this well. The questions are where the learning is.

The knowledge-telling trap

Teaching isn’t automatically good for the teacher. Rod Roscoe and Michelene Chi studied peer tutors and found two very different modes (Roscoe & Chi, 2007). In knowledge-telling, tutors recite what they already know — summarising the textbook, more or less fluently. It feels productive and teaches the tutor very little. In knowledge-building, tutors notice that they can’t answer something, work out why, connect it to what they do know, and extend their own understanding in the process. That is where tutors learn.

What tips a tutor from one mode to the other is usually a question they didn’t expect. Researchers building teachable agents — software students that learners teach — found that an agent that asked inquisitive follow-up questions moved learners into knowledge-building, and that they kept doing it after the prompts stopped (Shahriar & Matsuda, 2023).

There is a related finding from the same line of work: people try harder for someone they are teaching than they do for themselves. In studies with a teachable agent called Betty, students put more effort into learning so that their agent would do well — the protégé effect — and the effect was strongest for students who were otherwise struggling (Chase et al., 2009).

What a good student asks

So the listener matters a great deal, and the best listener is not an expert. An expert fills in your gaps without noticing, nods at your jargon, and corrects you before you have found the mistake yourself. A recent study of a teachable agent built on a large language model found exactly this weakness: the agent was too competent, which left the learner little to explain or correct (Chen et al., 2024).

A good student is curious and genuinely naive. They ask:

• Why? — Why does every word have to be compared with every other word?
• Can you give me an example? — the fastest way to find out whether you understand something or just its description.
• What does that word mean? — every term you use without defining is a place you might be hiding a gap.
• What would happen if…? — What if there were no position information? Questions like this test whether you understand the mechanism or have only memorised the outcome.

And they never tell you the answer. If they did, you’d stop doing the work that makes teaching useful.

Do it with a friend, or a rubber duck

You can do all of this without any special tool:

1. Pick one idea, not a whole subject. How a zip works, not mechanical engineering.
2. Close your notes. Teaching from memory is retrieval practice, and retrieval strengthens memory far more than rereading (Roediger & Karpicke, 2006).
3. Explain it to someone who doesn’t know it — a friend, a family member, or, if no one is around, a rubber duck or a blank page addressed to an imaginary newcomer.
4. Ask them, or yourself, the four questions above after each step. Keep going until one of them stops you.
5. Write each stopping point down as a question. Why don’t the teeth come apart under tension? That question is the most useful thing the session produced.
6. Afterwards, check your explanation against your notes or the source, not against how confident you felt. Mark what you said that the notes don’t support.
7. Go back and learn the gaps, then teach it again.

Two sessions a week on one idea will tell you more about what you understand than a month of rereading.

What “the Feynman technique” rests on

You may have seen this method called the Feynman technique: explain a concept in simple language as if to a child, find where you get stuck, go back to the source, simplify. The name is a modern popularisation. Richard Feynman was famous for explaining things clearly and for distrusting understanding that couldn’t be explained, but the four-step recipe attributed to him isn’t his.

The method doesn’t need his name. It rests on the research above: the illusion of explanatory depth, which says you can’t see your gaps until you try to explain; the evidence on learning by teaching, which says that explaining — especially in answer to questions — builds lasting understanding; and the evidence on retrieval, which says that doing it from memory is what makes it stick.

How Graphite does this

Graphite’s Teach it back is this exercise, built in. When you think you understand a note — a concept, or a project you’re working on — you close it and explain it to Graphite playing the Student: a curious newcomer who asks why, asks for an example, and asks what your words mean. It never corrects you and never reveals the answer. When you finish, Graphite steps out of the role and marks your explanation against your own notes — unsupported, question or link, never a score — and each gap goes under What’s missing in the note you taught, so you know exactly what to learn next. You can try it in your browser.