There is a sentence that gets passed around every study-skills video and every productivity thread: you remember 10% of what you read, 20% of what you hear, and 90% of what you teach. It is repeated with the confidence of a law of physics, usually attributed to a pyramid, a cone, or an institute.

The numbers are not real. They trace back to a set of unsourced percentages that appeared in a 1960s training manual, attached to a decades-old diagram of "experience" that never measured memory at all, and then hardened into folk knowledge. No study ever found the 90%. That matters, because the fake version of this idea is ruining a genuinely powerful technique: it promises too much, so when real results look modest, learners abandon a method that actually works.

Here is the honest version, in one paragraph. Teaching something you have just learned, or even just preparing to teach it, reliably improves your retention — by a small to moderate amount, under specific conditions, and for specific reasons that are well understood. It is one of the better-evidenced techniques in the learning literature. It is also easy to do badly. This article is about the evidence, the boundary conditions, and how to run the whole thing on your own with a language you are still learning.

What the Real Syntheses Show

The research program began, unusually, with a paper about teaching that was really about studying. In 1980, John Bargh and Yaacov Schul gave students material to learn with one instruction or the other: prepare to be tested, or prepare to teach it to someone else.

  • Bargh and Schul (1980), Journal of Educational Psychology 72(5), 593–604, DOI 10.1037/0022-0663.72.5.593. Students who prepared to teach outperformed those who prepared for a test on later recall and recognition — even though nobody actually taught anything yet.
  • Benware and Deci (1984), American Educational Research Journal 21(4), DOI 10.2307/1162999, replicated it with a real article on brain function and found the teaching-expectancy group did better on conceptual questions, not just rote recall — and reported higher intrinsic motivation while learning.

Then the effect was measured properly, in meta-analysis. Two syntheses matter most.

  • Kobayashi (2019), Japanese Psychological Research 61(3), 192–203, DOI 10.1111/jpr.12221, pooled 28 studies comparing preparing-to-teach against ordinary studying, and 16 studies comparing preparing plus actually teaching against ordinary studying.
  • Kobayashi (2024), Educational Psychology Review extended this to 39 studies comparing teaching against no teaching expectancy after preparatory learning.
The synthesis of 28 studies indicated that the estimated effect sizes (Hedges' g) were 0.35 for preparing-to-teach and 0.56 for teaching with preparing-to-teach. Both preparing-to-teach and teaching with preparing-to-teach were effective in promoting deep learning (as well as surface learning) and even after a delay. — Kobayashi 2019, abstract, Japanese Psychological Research

Read those two numbers carefully, because they contain the whole design of the technique. Preparing to teach — the intention, before you open the book — is worth roughly a third of a standard deviation, which is a real but unremarkable gain, comparable to many study-strategy tweaks. Adding the act of explaining out loud roughly doubles it. If you want the bigger effect, you have to open your mouth. This site's other articles on the testing effect and on desirable difficulties describe the same shape from different angles: retrieval and generation are what pay, and passive review is what feels like work without being it.

Two more figures to keep the picture calibrated:

  • Cohen, Kulik and Kulik (1982), American Educational Research Journal 19(2), 237–248, DOI 10.3102/00028312019002237, synthesized 38 peer-tutoring programs and found that the tutors — not the tutees — gained about 0.33 standard deviations in achievement. This is the finding that justifies the technique for language exchange partners who are giving away their own time.
  • Ribosa and Duran (2022), Educational Research Review 37, 100475, DOI 10.1016/j.edurev.2022.100475, pooled the newer strand where students create teaching materials rather than lecture, and found an overall effect of only 0.17 standard deviations — stable across school levels and subjects, but small.

And this one, from the same family, is the most useful single sentence in the literature for anyone designing a routine:

A learning-by-teaching effect can also occur when students provide explanations to a real, remote, or even fictitious audience that cannot be interacted with. It is unclear, however, which underlying mechanisms drive learning by non-interactive teaching effects and why several recent studies did not replicate this effect. — Lachner, Hoogerheide, van Gog and Renkl (2022), Educational Psychology Review 34, 575–607, DOI 10.1007/s10648-021-09643-4

An imaginary audience works. That is the sentence that makes this technique available to a learner who has nobody to teach. It also carries the warning that follows: the effect is fragile if you do the wrong kind of explaining.

Teaching Is Three Different Activities Wearing One Name

Most of the confusion in this literature — and the reason some studies find nothing — comes from lumping together three stages that Bargh and Schul separated in 1980 and that later work has tested individually.

  • Preparing to teach. You know that you will have to explain this, so you read differently: you look for the shape of the idea rather than the sentence in front of you. This changes how you organize material at intake.
  • Explaining to others. You have to put the idea into your own words, in order, out loud, which forces elaboration and exposes the gaps where you can say the term but not use it.
  • Interacting. Someone asks a question you cannot answer, and for the first time you notice precisely where your understanding stops.

The cleanest experiment on the split is Fiorella and Mayer (2013), Contemporary Educational Psychology 38(4), 281–288, DOI 10.1016/j.cedpsych.2013.06.001. Learners studied a lesson on the Doppler effect under either a test expectancy or a teaching expectancy. Those who expected to teach — without ever teaching — beat the test group on an immediate comprehension test, by about d = 0.59. But when the same comparison was made after a one-week delay, the advantage had faded (d = 0.24, not statistically significant).

The follow-up, Fiorella and Mayer (2014), Contemporary Educational Psychology 39(2), 75–85, DOI 10.1016/j.cedpsych.2014.01.001, added the missing piece. In their second experiment, students either prepared to teach and then actually delivered a short recorded lecture, or prepared to teach and simply restudied. On a delayed test, only the group that had both prepared and taught showed the benefit — about d = 0.56. The abstract states the conclusion bluntly: there was a teaching-expectancy effect for short-term learning, and preparing to teach coupled with actually teaching produced the greatest long-term learning.

For a language learner, that is a design brief. Reading the grammar chapter with a vague intention to learn it is stage zero. Deciding that you will explain the chapter to somebody on Thursday is stage one. Standing up and doing it is stage two. Only the last one survives the week.

Why It Works: Generative Processing, Not Charisma

The accepted account is not motivational, and it is not about "learning styles" or the alleged power of teaching as a social role. It is about what psychologists call generative processing: building a mental model rather than filing a sentence.

Roscoe and Chi (2007), Review of Educational Research 77(4), 534–574, DOI 10.3102/0034654307309920, studied what actually separates tutors who learn from tutoring from those who do not. Their distinction is the most practically useful idea in this whole area: knowledge-telling versus knowledge-building. A tutor who recites a summary — reads the definition back, lists the steps in the order of the textbook — learns very little. A tutor who is forced to build an explanation — to connect the new rule to something else, to find an example that is not in the book, to answer "but why is it like that?" — learns a great deal. The activity looks identical from outside. The cognitive work is completely different.

This connects to a second research line: Chi, de Leeuw, Chiu and LaVancher (1994), Cognitive Science 18(3), 439–477, DOI 10.1016/0364-0213(94)90016-7, showed that simply prompting learners to explain a text to themselves as they read — self-explanation — produced large gains over reading without prompts. Fiorella and Mayer (2016), Educational Psychology Review 28, 717–741, DOI 10.1007/s10648-015-9348-9, later placed learning by teaching among eight "generative learning" strategies, alongside self-explanation, summarizing, mapping, drawing and enacting — exactly the family this site has written about before.

There is also a motivational component that the research is unusually clear about, because it was measured directly. Chase, Chin, Oppezzo and Schwartz (2009), Journal of Science Education and Technology 18(4), 334–352, DOI 10.1007/s10956-009-9180-4, had students build a "teachable agent" — a simple animated character they had to instruct. Students who believed they were teaching someone else worked measurably harder and longer than students learning the same material for themselves. They gave the name the protégé effect to this willingness to put in extra effort on behalf of a learner. The effort is the mechanism; the teaching is the delivery system.

When It Backfires

Three conditions reliably erode or reverse the effect. Knowing them is what separates a working routine from a waste of twenty minutes.

1. When the audience makes you panic, not prepare

Renkl (1995), Learning and Instruction 5(1), 21–36, DOI 10.1016/0959-4752(94)00015-h, had students prepare to teach probability while working through examples, and found no teaching-expectancy benefit at all. The most plausible explanation in the literature is stress: anticipating a real performance can consume the working memory the task needs. If your version of "prepare to teach" is "prepare to be judged live in front of people who are better than me", you have imported the anxiety research this site has covered elsewhere — and anxiety is a language learner's single biggest reported barrier.

2. When you tell instead of build

The Roscoe and Chi distinction again. Reciting is not teaching. If your explanation could be produced by reading the page aloud, no generative work is happening, and the meta-analytic effect collapses toward the small numbers in the material-creation studies.

3. When the modality is wrong

Lachner and colleagues' model compares spoken and written explanations and finds spoken non-interactive teaching generally outperforms written — video-recorded or spoken explanations beat writing out an explanation. A cumulating meta-analysis of non-interactive teaching put the spoken-and-written pool at about g = 0.22. And a 2025 classroom study of 317 secondary-school students, published in Educational Psychology Review, found that teaching beat restudying on immediate conceptual knowledge but produced no lasting benefit after a delay, and left students under-confident about what they had learned.

Notice the pattern in the failures: no audience, no speech, no effort to build. Those are exactly the conditions produced by "I'll just summarise the chapter in my head".

What This Means for a Language Learner Specifically

Almost all of the meta-analytic evidence above comes from science and history content, not from second-language learning. That is worth stating plainly rather than pretending otherwise. But the language-learning strand does exist, and it points the same way.

Swain, Brooks and Tocalli-Beller (2002), Annual Review of Applied Linguistics 22, 171–185, DOI 10.1017/s0267190502000090, reviewed the evidence that peer-peer dialogue is itself a means of second-language learning — not just practice, but knowledge construction. In their account, learners who explain a grammar point to a partner, negotiate why a form is wrong, and repair their own explanation are doing the cognitive work that produces durable knowledge. This is the second-language version of the knowledge-building condition, and it is why group work sometimes teaches the explainer far more than the explained.

Three specific consequences follow for someone studying a language with flashcards.

  • Explaining grammar in the target language is the double win. You are generating output (which this site has argued is non-optional), and you are reorganising the rule on the way out. Explaining a rule you can only half-use is where the gap becomes visible.
  • In a language exchange, the tutor role is where the measurable gain is. Cohen's 1982 synthesis found the tutors gained about 0.33 SD. If you swap languages with a partner, the hour you spend teaching your own language to someone else is not charity — it is the better-evidenced half of the exchange.
  • The audience does not have to exist, and the language does not have to be perfect. Lachner's finding that a fictitious audience works is the licence to do this alone. If your productive vocabulary will not stretch to a full explanation, explain in the target language and allow yourself the source language for the meta-commentary — with one rule: the sentence you write on the card must be in the target language.

The 20-Minute Teaching Loop: A Seven-Step Protocol

This is designed for one person, once a week, with no partner and no camera if you do not want one. It fits around a reviewing session of the kind this app already produces.

Step 1 — Pick the thing you cannot do

Not the chapter you just read comfortably. The three to five flashcards you failed most often this week, one grammar rule that keeps collapsing mid-sentence, or one pronunciation contrast you still cannot hear. Teaching something you already know cold produces no generative work.

Step 2 — Set an empty chair and give it a name

Refuse to let it be abstract. "Explain the subjunctive to Mina, who has never seen it" is a different task from "review the subjunctive". The fictitious learner only works if you specify what they do and do not know — which forces you to choose what to leave out, an act of organization in itself.

Step 3 — Write the lesson plan before you open the book to teach

Ten minutes, three lines, on paper. This is the preparing-to-teach stage, worth the g = 0.35 on its own. A workable shape:

1. What they will be able to DO at the end
2. The one rule I will state (in one sentence)
3. Two examples — one correct, one wrong, with what changes

Step 4 — Teach it out loud, standing, for three minutes

Spoken, not written — Lachner's modality result. Standing matters less for the physics than for the honesty: it is very hard to mumble through a gap while standing up and pretending someone is listening. No notes after the first thirty seconds. When you stumble, do not stop and go look it up. Mark the spot on your paper with a single dot and carry on.

Step 5 — Collect the dots

Every dot is a card that was lying to you. That, not the lesson, is the actual product of the exercise: teaching for three minutes tells you exactly which cards to rewrite, which mnemonic to replace, and which rule you only recognize rather than use.

Step 6 — Answer the imaginary question

Ask the hardest question you can invent — "but why doesn't it apply to this word I already know?" — and answer it aloud, in full sentences. This is the interactive stage reproduced alone, and it is where knowledge-building happens. If you cannot answer, you have found next week's lesson for free.

Step 7 — Two minutes, same day next week

Re-teach the same three items seven days later without preparation. The Fiorella and Mayer delayed test is the one that separates the techniques that work from the ones that only feel like work. If the explanation comes out clean and unhesitating a week later, the material has moved. If it collapses, the loop is doing its job by telling you so.

The Rules That Keep It Honest

  • Teach at the edge, not the center. Anything you can already explain fluently has stopped being generative.
  • Build, do not recite. If your explanation is the textbook's sentences in the textbook's order, you are doing stage one at best. Invent one new example every time.
  • Speak it. Written explanations still help, but spoken explanations carry the larger effect.
  • One concept per session. The generative account is about working memory: three overlapping rules explained badly is a worse routine than one rule explained until it holds.
  • Keep the dots. A teaching session with no note of where you faltered is a performance, not a study session.
  • Expect modest gains and take them. The honest effect is around half a standard deviation when you do it properly and much less when you do it as a warm-up. The fake 90% figure would have you discard a technique that is genuinely in the top tier of things you can do with twenty minutes.

The Takeaway

The research on learning by teaching is, in the end, unusually flattering to learners without resources. The technique requires no app, no teacher and no partner — only the willingness to stand up and explain something to a chair, out loud, badly, and then write down exactly where it went wrong. The person who benefits most is the explainer. The person who benefits second most is the one who writes down the dots.

For a language learner, that turns the weekly review into something more than review. Take the cards that fought back, teach them to a named imaginary student, and let your own stumbles tell you what to fix next. It is the only study technique on this site where the measure of success is the fact that you embarrassed yourself a little.