What Is the Protege Effect?
The protege effect (also spelled "protégé effect") is the finding that people learn material more thoroughly when they prepare to teach it to someone else than when they study it only for themselves. The term was coined by Catherine Chase, Doris Chin, Marily Oppezzo, and Daniel Schwartz in 2009, who wrote that students "make greater effort to learn for their TAs than they do for themselves." Pooled across dozens of experiments the gain is moderate, roughly g = 0.17 to 0.56 depending on which synthesis you read, and it depends heavily on knowing in advance that you will teach.
The mechanism is not mysterious. When you know someone else will rely on your explanation, you study differently: you organize more carefully and you go hunting for the gaps, because you have to anticipate where a reader will get confused. Passive reading rarely reaches that level.
Most people sense this intuitively. Think about any time you've explained a concept to a colleague. The act of explaining forced you to confront fuzzy spots in your own understanding. That discomfort is the protege effect at work.
What makes it distinct from other learning strategies is the social dimension. It's not just about reformulating knowledge, which is closer to the Feynman Technique. It's about reformulating knowledge for someone else. Whether that social layer adds anything beyond the reformulation is one of the genuinely open questions in this literature, and the section on limits below is where the honest answer lives.
What the Research Shows: Studies and Meta-Analyses
Chase et al. (2009): The Original "Teachable Agents" Study
Chase and colleagues published "Teachable Agents and the Protege Effect: Increasing the Effort Towards Learning" in the Journal of Science Education and Technology. The study used Betty's Brain, a system where students taught a virtual agent by building concept maps about science topics.
Students who believed they were teaching Betty spent more time on learning activities and scored higher than students using the same system for themselves. Worth knowing about the design: in the first study both groups ran the same software and differed only in what they were told, and the sample was 62 eighth-graders. A second study replicated the time-on-task effect with 24 fifth-graders and used think-aloud protocols to probe why. It's a real finding with a small base, not a settled quantity.
Nestojko et al. (2014): Expectancy Alone Can Be Enough
John Nestojko, Dung Bui, Nate Kornell, and Elizabeth Bjork isolated the teaching-expectancy variable. One group studied a passage expecting a test; the other expected to teach it. Nobody actually taught.
In Experiment 1 the teach-expectancy group recalled more of the passage (d = 0.56), recalled it more efficiently (d = 0.79), clustered their recall more closely to the passage's own paragraph structure (d = 0.65), and scored higher on a short-answer test. Experiment 2 is the part usually left out: there the main effect of expectancy was not significant (p = .389), and the advantage on the passage's main points was only marginal (p = .083). One influential experiment, one that mostly did not replicate it, in the same paper.
Kobayashi (2019, 2024): By How Much?
Individual studies show a direction; meta-analyses tell you the size. Keiichi Kobayashi's 2019 synthesis in Japanese Psychological Research pooled 28 studies and found g = 0.35 for preparing to teach and g = 0.56 for preparing and then actually teaching, both against studying with no teaching expectancy.
His 2024 meta-analysis in Educational Psychology Review pooled 39 studies and reported g = 0.27 overall, and it found the moderator that matters more than any other in this article. Teaching after studying with a prior expectation of teaching gave g = 0.48. Teaching after studying without that expectation gave g = -0.02, statistically indistinguishable from nothing. Preparation, not performance, is doing the work.
A separate 2019 paper of his in Frontiers in Psychology looked at interactivity, comparing face-to-face teaching against indirect formats like a recorded lecture or a written explanation. Expecting to teach face-to-face pooled at g = 0.50 against g = 0.27 for indirect, and actually going through with it reached g = 0.84 for direct teaching and g = 0.48 for indirect. Treat the 0.84 with care: it rests on four group comparisons, and Kobayashi himself calls the evidence "still inadequate and subject to some exceptions."
Ribosa and Duran (2022): The Lowest and Most Conservative Estimate
Jesus Ribosa and David Duran pooled 23 studies and 62 comparisons in Educational Research Review and landed at g = 0.17, with no publication bias detected. Two qualifications matter. In their studies learners generated teaching materials, things like video lectures, test questions and instructional texts, rather than taught a person, which Kobayashi calls an overlapping but distinct category. And the benefit shows up when that is compared against interventions not expected to help, rather than against other interventions that were. Measured against a genuinely active control, the advantage is not statistically significant.
Koh, Lee, and Lim (2018): It's the Retrieval That Does the Work
Aloysius Koh, Sze Chi Lee, and Stephen Lim tested what part of teaching drives the benefit, in Applied Cognitive Psychology. They compared four groups: teaching from memory, teaching while reading from prepared notes, retrieving the material without teaching, and an arithmetic control.
On a comprehension test one week later, teaching from memory and pure retrieval practice both beat the other two, while teaching with notes in hand was no better than the control. Close the book, then explain. If you teach with the source open, you skip the retrieval and forfeit the benefit. Note the other half of that result: plain retrieval practice, with no teaching at all, did just as well.
The Evidence at a Glance
| Synthesis | Studies pooled | Pooled estimate | The catch |
|---|---|---|---|
| Kobayashi (2019), Japanese Psych. Research | 28 studies | g = 0.35 preparing, 0.56 preparing and teaching | Publication bias assessed only with Rosenthal's fail-safe N, a method now widely criticized |
| Kobayashi (2019), Frontiers in Psychology | 44 group comparisons | g = 0.50 direct vs 0.27 indirect expectancy | The headline 0.84 for direct teaching rests on k = 4 |
| Kobayashi (2024), Educational Psych. Review | 39 studies | g = 0.27 overall; 0.48 with expectancy, -0.02 without | Heterogeneity I-squared of 76 to 94%, so the single pooled figure means little |
| Ribosa and Duran (2022), Educational Research Review | 23 studies, 62 comparisons | g = 0.17 | Not significant against an active control; no publication bias detected |
| Bisra et al. (2018), Educational Psych. Review | 69 effect sizes | g = 0.55 for self-explanation | The solo, audience-free version scores at least as high as the social one |
| Chang et al. (2025), Educational Psych. Review | 32 studies | g = 0.39 for tutors in cross-age tutoring | No moderation by session count, tutor type, or subject area |
Limits and Moderators: When Teaching Doesn't Help
Every genuine review of this literature has to report the failures, and there are enough of them to change how you use the technique.
Expecting to teach does not reliably survive a delay. Fiorella and Mayer (2013) taught undergraduates about the Doppler effect. Students who prepared to teach beat the control on an immediate comprehension test, but on a one-week delayed test, expecting to teach produced no benefit over the control. Their 2014 follow-up found that students who actually taught were the ones who performed best at delay.
Renkl (1995) found no benefit at all from preparing to teach probability worked examples. Guerrero and Wiley (2021) is the strongest paper on the other side, with about 210 participants per experiment, but its effects are small (partial eta-squared of .02 in both) and its delayed claim rests on a non-significant interaction. Lachner and colleagues, reviewing the area in Educational Psychology Review in 2022, open by asking why several recent studies "did not replicate this effect."
A live audience is not clearly better than an imagined one. This cuts against the intuitive story. Wang, Lin, and Chen (2021) ran 597 secondary students through teaching expectancy crossed with peer teaching, imagined teaching, and no teaching. Imagined teaching improved both immediate and delayed comprehension; peer teaching improved only immediate. The authors advise caution either way. Kobayashi's interactivity analysis points the opposite direction. The picture is unresolved, which means anyone telling you a real audience is required is ahead of the evidence.
Several well-run studies find nothing. Jacob, Lachner, and Scheiter (2021) compared explaining aloud, writing in an editor, writing in a messenger chat, and a retrieval control across 137 participants: no differences in learning. What did differ is instructive. Explaining was rated more effortful, more enjoyable, and more interesting than retrieving, while producing the same result. Effort and enjoyment are not evidence of learning. Lachner, Jacob, and Hoogerheide (2021) found no effect of explaining at all in one experiment (N = 147), and in a second (N = 51) self-explaining beat explaining to a fictitious student.
Novice teachers tend to tell rather than explain. Roscoe and Chi (2007), in the Review of Educational Research, documented what they called a knowledge-telling bias: tutors "simply revealed answers, summarized facts, or described procedures with little elaboration," and novice tutors are poor judges of their own understanding of what they teach. The cognitive benefit comes from the elaboration, and the default behavior skips it.
One documented harm, on motor skills. Daou and colleagues (2019) had people learn golf putting either expecting to teach it or expecting a test. The teach group performed better under low pressure and then fell back to the test group's level under high pressure. A registered-report replication with 156 participants confirmed it. If you're learning something you'll have to execute under pressure, preparing to teach it may not be neutral.
What does not seem to matter: learner age and subject area. Ribosa and Duran describe the effect as seeming stable across educational levels and content areas, and Chang et al. (2025) found no moderation by number of sessions, tutor type, or subject in 32 cross-age tutoring studies.
So the defensible claim is narrower than the popular one. Committing in advance to explain something, then explaining it from memory, is a good use of study time. It is not reliably better than simply testing yourself, and the social element is doing less work than the story suggests.
Why Teaching Works: The Cognitive Mechanisms
Teaching activates several processes at once, each of which independently improves learning. Which of them is actually responsible is still contested.
Retrieval Practice
Koh, Lee, and Lim proposed the retrieval practice hypothesis: teaching helps because it forces you to pull material out of memory, and the teaching itself adds little. Their notes-in-hand condition, which removes the retrieval and keeps the teaching, is the clean test and it showed no benefit.
Be careful how strongly you state this. Kobayashi assessed the hypothesis in Frontiers in Psychology in 2022 and concluded that "the currently available evidence is inadequate to assess the hypothesis." Lachner and colleagues' 2022 review treats retrieval as one candidate mechanism among several, not the established answer.
Metacognitive Monitoring
Teaching forces you to evaluate your own understanding in real time. As you explain, you're checking whether it holds together. This kind of self-monitoring is central to self-regulated learning (Dunlosky and Metcalfe, 2009).
When you study for yourself you can gloss over fuzzy understanding. When you study to teach, fuzziness becomes a problem you have to solve.
Elaborative Retrieval
Preparing to teach makes you reconstruct reasoning, generate examples, and build analogies rather than just recall facts, the family of activities Fiorella and Mayer (2016) grouped under generative learning. One caution on the evidence usually cited here: Karpicke and Blunt (2011) showed in Science that retrieval practice beat elaborative studying with concept mapping, and their own conclusion cuts against the elaboration framing, attributing the gain to "retrieval-specific mechanisms rather than by elaborative study processes."
Responsibility, With a Caveat
The popular account says another person's dependence on you changes the stakes, and Biswas, Leelawong, Schwartz, Vye, and the Teachable Agents Group at Vanderbilt (2005) treated responsibility as a design principle for teachable agents. Two honest qualifications. The phrase "responsibility effect" does not appear in that paper and is not an established term in this literature. And Chase et al. cite the observation as anecdotal evidence of students developing a feeling of responsibility toward their agents, which is how it entered the literature.
The measured version of the claim is weaker and more interesting. Kobayashi's interactivity analysis is consistent with a real audience helping; Wang, Lin, and Chen's 597-student experiment points the other way. Treat responsibility as a plausible mechanism, not a demonstrated one.
Organizational Processing
To teach something you have to decide what comes first, what depends on what, and what can be skipped. Nestojko et al. observed a version of this: participants expecting to teach recalled the passage in an order that tracked its paragraph structure more closely than the test-expectancy group did. Note what that is and isn't. It's evidence of restructuring toward the source's organization, not of building a deeper hierarchy, and a second organization measure in the same experiment showed no difference.
The Learning Pyramid: Why the 90% Number Is a Myth
You've probably seen the Learning Pyramid: a tidy triangle assigning each activity a retention percentage, with "teaching others" crowned at 90% and reading at 10%. It's usually credited to the National Training Laboratories in Bethel, Maine. Here is the version that circulates:
| Learning Activity | Claimed Retention Rate |
|---|---|
| Lecture (passive listening) | 5% |
| Reading | 10% |
| Audio-Visual | 20% |
| Demonstration | 30% |
| Discussion Group | 50% |
| Practice by Doing | 75% |
| Teaching Others | 90% |
Those numbers are not real. Kåre Letrud and Sigbjørn Hernes traced the pyramid's history in Cogent Education (2018) and found that versions of it have been circulating in educational debate for more than 160 years, concluding that the models "did not originate from empirical research." Asked directly about the figures, NTL replied: "Yes, we believe it to be accurate, but no, we no longer have, nor can we find, the original research that supports the numbers" (quoted in Lalley and Miller, 2007). Note the precise claim. NTL said the research cannot be located, not that it was never done. Either way there is nothing to check.
The suspiciously round, evenly spaced figures are the giveaway. Real memory data never lines up at 10, 20, 30, 50.
So why show the pyramid? Because the ranking it implies, that active production beats passive consumption, is broadly supported even though the percentages are fiction (Dunlosky et al., 2013; Karpicke and Blunt, 2011). State the benefit as an effect size of roughly 0.17 to 0.56, not as 90%. If you see the pyramid cited as hard data, treat it as a signal about the source.
Study-for-Self vs. Study-to-Teach: A Direct Comparison
What changes when you shift from "I'm studying this for myself" to "I'm studying this to teach someone else"?
| Dimension | Study for Self | Study to Teach |
|---|---|---|
| Information organization | Follows source order loosely | Restructured; recall tracks the source's own structure more closely |
| Focus | Broad coverage, tries to absorb everything | Selective focus on key concepts and relationships |
| Gap detection | Low; easy to skip over confusion | High; gaps become obstacles to a clear explanation |
| Retention after 1 week | Baseline | Higher only if you committed in advance and explain from memory |
| Depth of understanding | Surface to moderate | Deeper, when elaboration actually happens rather than knowledge-telling |
| Motivation | Extrinsic (pass the test) | Prosocial (help someone else understand) |
| Study strategy | Re-reading, highlighting, note-taking | Summarizing, generating examples, creating analogies |
| Error correction | Often deferred or ignored | Addressed immediately; errors would confuse the learner |
The study-to-teach mindset doesn't demand more time. It demands a different intention, formed before you start reading. That ordering is not a stylistic preference: it's the difference between g = 0.48 and g = -0.02 in Kobayashi's 2024 data.
The Protege Effect vs. the Feynman Technique
The Feynman Technique and the protege effect are related but distinct, and the comparison is less lopsided than it's usually presented.
The Feynman Technique asks you to explain a concept in simple language as if teaching a beginner: choose a concept, explain it plainly, find the gaps, simplify again. As normally practiced it's a solo exercise, with no real person depending on the result.
The protege effect adds the social layer. But here's the finding that should temper the usual hierarchy: self-explanation, the solo cousin, pools at g = 0.55 across 69 effect sizes (Bisra et al., 2018), which is at the top of the range for learning by teaching and above most estimates of it. The audience may be adding less than the framing implies.
| Feature | Feynman Technique | Protege Effect |
|---|---|---|
| Audience | Imagined or self | Real person or public audience |
| Accountability | Internal only | Social; someone depends on your accuracy |
| Primary mechanism | Self-explanation, gap detection | Retrieval plus metacognitive monitoring |
| Best for | Deep understanding of single concepts | Sustained motivation and a reason to finish |
| Evidence base | Pools at g = 0.55 for self-explanation | Pools at g = 0.17 to 0.56, with expectancy as the key moderator |
| Social component | Optional | Present, but its added value is contested |
The practical read: do both, and don't assume the audience is what's working. The two things both methods share, retrieval from memory and forced elaboration, are the parts with the strongest support. A real audience is best understood as a commitment device, which is not a small thing, because learning in public is what makes you actually do the explaining.
Does the Protege Effect Work When Your Student Is an AI?
Since the original teachable agents were scripted, an obvious question is whether a large language model makes a better protege. The 2024 to 2026 evidence says: not yet, and the reason is specific.
The cleanest test is a null. Xu, Zhang, Tang, and Lee (2026) randomized 96 participants across four conditions: explaining to an LLM playing a tutee, a peer, a challenger, or a near-passive control. On the post-test, no condition beat the control (F(3,91) = 1.54, p = .21). Effort clearly rose, with the tutee group writing 353.8 words on average against 121.3 for the control. It just didn't convert into knowledge. The tutee group also reported the most pressure of the three agent roles, 4.02 against 2.49 for the peer and 2.93 for the challenger, and no better than the passive control at 4.06. That is the opposite of the ego-protective buffer Chase et al. (2009) described, where teaching an agent was supposed to shield the learner from the sting of being wrong.
A widely cited paper names the defect and never measures learning. Jin, Lee, Shin, and Kim (CHI 2024) built TeachYou and its LLM tutee AlgoBo, and reported that "LLMs' expansive knowledge as tutees discourages learners from teaching." Their name for the failure mode is under-teaching: learners stop trying because the agent already knows. The paper notes as a limitation that it ran no pre-post test, and its metacognition intervention showed no significant gain.
Positive results exist and each carries a catch. Chen, Wei, Le, and Zhang (2025) found that students who taught a ChatGPT agent scored higher on a programming knowledge test (n = 41), but code correctness was not significantly different, and the authors' own explanation is telling: ChatGPT tends to produce correct code, which removes the debugging practice. The largest deployment, Wang and colleagues' semester-long study at EPFL (2026) covering 453 analyzed students and 3,809 dialogues, found the teaching group needed about 4.9 percent fewer quiz attempts than a reading baseline. Note what the outcome is: quiz attempts, not a knowledge test, and it cost 14.4 minutes of work against 3.4. Around a quarter of dialogues contained pasted-in external content, and that predicted worse performance.
Believability isn't the obstacle. Rogers and colleagues (2025) ran 119 students across four experiments, one of which compared an LLM tutee against a human confederate pretending to be a novice. Perceived intelligence was not significantly different. Students believe the act.
The model won't stay a novice. This is the real blocker. Song, Guo, and Lin (2026) put it directly: even when prompted to act like a novice, LLMs "can still produce expert-level explanations," drifting beyond the intended knowledge level. Worse, they capitulate. Do, Sonkar, and Sachan (2026, a preprint) measured how often simulated students abandon a misconception and found the smallest model they tested, Qwen3-4B, flipping its answer 92 percent of the time on feedback that says nothing more than "you're wrong." A tutee that agrees with whatever you say cannot show you that your explanation failed, and that failure signal is the point.
There's also a citable hole in this literature: nobody has replicated the original responsibility manipulation with an LLM tutee. The LLM teachable-agent papers lean on Chase et al. for the mechanism without testing it.
What to do with that. Use an LLM as a rehearsal partner rather than a protege. Explain from memory with the source closed, which is the part with the best evidence, then ask it to find the gaps in what you said. Instruct it to hold a wrong position until you genuinely correct it, and treat easy agreement as a sign the exercise is not working. If you want the protege effect proper, the evidence still points to a person.
Five Ways to "Teach" Without Being a Teacher
You don't need a classroom or students to activate the protege effect. What you do need, based on everything above, is to decide before you read that you'll explain this to someone, and then explain it without looking.
1. Make Your Highlights Public
When you highlight a passage using Glasp's web highlighter and add a note explaining why it matters, you're performing a micro-teaching act. Your highlights appear on your public profile where other learners can find them. That visibility shifts your processing from "this is interesting to me" to "this needs to be clear enough for someone else."
The ordering matters more than the tool. Decide you're annotating for a reader before you start the article, not after. Over time the collection becomes a teaching resource: a curated reading list with context others can follow.
2. Write Blog Posts and Summaries
Writing a summary for an audience forces the same organizational and elaborative work formal teaching does. You have to decide what's essential, build a logical flow, and put ideas in your own words. The audience doesn't need to be large.
Using YouTube Summary to pull key points out of lectures and talks gives you raw material. The learning happens when you close it and write the explanation yourself.
3. Join or Create Study Groups
Study groups with teaching rotations activate the effect systematically; groups where everyone reads silently don't. Assign specific concepts in advance so each member knows days ahead that they're responsible for one. Have each person teach without notes while others ask questions. The questions do real work, because they force the teacher to elaborate rather than recite.
4. Practice Explaining to AI, With Realistic Expectations
Glasp's AI chat lets you rehearse an explanation without recruiting a human. Be clear-eyed about what this buys you. As the section above lays out, the randomized evidence on teaching an LLM tutee is currently null, and the value you can reliably get is retrieval and gap-finding, not the social mechanism.
So use it that way. Close the source, explain the concept from memory, then ask the AI to identify what you left out, what you got wrong, and what a confused reader would still ask. Push back when it agrees too readily. You're using it as an examiner, not a student.
5. Share on Community Platforms
Posting your learning on the community feed puts your understanding in front of real people. When someone comments, asks a question, or offers a different reading, you get feedback solo study can't produce. Whether a live audience adds anything beyond the explaining itself is unsettled, as the limits section says. What it reliably does is make you follow through.
Learning in Public: The Protege Effect at Scale
Learning in public is the protege effect applied to your whole practice. Instead of studying privately and occasionally teaching, you share what you read as a matter of course, which turns the expectancy from something you have to remember to summon into the standing condition of how you read.
That's the strongest argument for it. The single best-supported moderator in this literature is anticipating an audience before you study, and a habit of sharing means you never study without one.
Glasp was built around this. When you export your highlights into a blog post, share them, or leave them visible on your profile, the annotation has to make sense to somebody who wasn't there. Pair that with active recall and you're exercising the two mechanisms with the best evidence behind them.
The benefits go past individual retention. Your public learning becomes a contribution to collective intelligence. Others build on your highlights; you find new angles through theirs.
Frequently Asked Questions
What is the protege effect in simple terms?
The protege effect is the finding that you learn material better when you prepare to teach it to someone else than when you study it only for yourself. Preparing to teach pushes you to organize the material, notice your own gaps, and retrieve it from memory. Nestojko et al. (2014) showed that even expecting to teach, without ever teaching, improved recall in one experiment, though a second experiment in the same paper found no overall effect.
By how much does teaching others improve learning?
The honest answer is a range of effect sizes, not a percentage. Kobayashi's 2019 meta-analysis of 28 studies found g = 0.35 for preparing to teach and g = 0.56 for preparing and teaching. His 2024 meta-analysis of 39 studies found g = 0.27 overall. The most conservative synthesis, Ribosa and Duran (2022), found g = 0.17 and no significant advantage over an active control. Ignore the "90% retention" figure from the Learning Pyramid; Letrud and Hernes (2018) showed it has no empirical source.
What single thing most determines whether it works?
The deciding factor is whether you knew in advance that you would teach. In Kobayashi's 2024 meta-analysis, teaching after studying with a prior teaching expectancy gave g = 0.48; teaching after studying without it gave g = -0.02. Decide before you open the material, not after.
Is teaching better than just testing yourself?
Often not, and that is the most under-reported result in this literature. Koh, Lee, and Lim (2018) found that pure retrieval practice matched teaching from memory on a comprehension test a week later, and both beat teaching with notes in hand. Self-explanation pools at g = 0.55 (Bisra et al., 2018), higher than most estimates for learning by teaching. If your goal is retention alone, active recall is the cheaper route. Teaching adds motivation, an audience, and a reason to finish.
How is the protege effect different from the Feynman Technique?
The Feynman Technique is about explaining a concept in simple language to find your knowledge gaps, usually alone. The protege effect adds a real or anticipated audience. The mechanisms they share, retrieval and elaboration, have the strongest evidence; how much the audience itself contributes is still contested, with Wang, Lin, and Chen (2021) finding that imagined teaching improved delayed comprehension while actual peer teaching did not.
Does the protege effect work if you teach an AI?
On current evidence, not reliably. The cleanest randomized test (Xu et al., 2026, N = 96) found no learning advantage for explaining to an LLM tutee compared with a near-passive control, even though participants clearly worked harder. The main reason is that language models won't stay novices: they drift into expert-level explanations and accept whatever correction you offer. Use an AI to quiz you on an explanation you gave from memory, which is the part with evidence behind it, rather than as a student you're teaching.
Do I need to be an expert to benefit from teaching others?
No, and there's a real risk in the other direction. Chase et al. (2009) found learning gains in middle schoolers who were experts in nothing. But Roscoe and Chi (2007) documented that novice tutors default to knowledge-telling, reciting facts with little elaboration, and judge their own understanding poorly. The benefit lives in the elaboration, so push yourself to explain why, not just what.
How does the protege effect relate to active recall?
Teaching from memory is a form of active recall: you retrieve, you elaborate, and you monitor your accuracy. That overlap is probably most of the effect. Koh, Lee, and Lim's retrieval practice hypothesis makes it explicit, though Kobayashi's 2022 review concluded the evidence isn't yet strong enough to settle it.
Conclusion: What the Evidence Actually Supports
Learning by teaching works, and it works for less mystical reasons than the popular version claims. Pooled across the good studies it's a modest-to-moderate gain, somewhere between g = 0.17 and g = 0.56, and almost all of it depends on deciding in advance that you'll have to explain this to someone. Teach from your notes and you get nothing. Get told to teach after you've studied and you get nothing. Commit first, then explain from memory, and you get a real effect that overlaps heavily with plain retrieval practice.
That's a narrower claim than "teaching others gives you 90% retention," and it's more useful, because it tells you exactly which part to keep.
The practical version is short. Before you read something, decide who you're going to explain it to. Highlight with Glasp's web highlighter as if you're annotating for that person. Close the tab and explain it from memory, to a study group, a blog audience, or the community feed. Use Glasp's AI chat to interrogate the explanation you just gave rather than to teach a student who already knows the answer. Turn a YouTube Summary into your own written argument instead of a transcript.
The commitment comes first. Everything the research supports follows from it.
References
Bisra, K., Liu, Q., Nesbit, J. C., Salimi, F., and Winne, P. H. (2018). Inducing self-explanation: A meta-analysis. Educational Psychology Review, 30, 703-725. doi:10.1007/s10648-018-9434-x
Biswas, G., Leelawong, K., Schwartz, D., Vye, N., and The Teachable Agents Group at Vanderbilt (2005). Learning by teaching: A new agent paradigm for educational software. Applied Artificial Intelligence, 19(3-4), 363-392. doi:10.1080/08839510590910200
Chang, A., Mauer, E., Wanzek, J., Kim, S., Scammacca, N., and Swanson, E. (2025). Examining the academic effects of cross-age tutoring: A meta-analysis. Educational Psychology Review, 37. doi:10.1007/s10648-025-09997-z
Chase, C. C., Chin, D. B., Oppezzo, M. A., and Schwartz, D. L. (2009). Teachable agents and the protege effect: Increasing the effort towards learning. Journal of Science Education and Technology, 18(4), 334-352. doi:10.1007/s10956-009-9180-4
Chen, A., Wei, Y., Le, H., and Zhang, Y. (2025). Learning by teaching with ChatGPT: The effect of teachable ChatGPT agent on programming education. British Journal of Educational Technology, 57(1), 163-184. doi:10.1111/bjet.70001
Cabral, D. A. R., Daou, M., Bacelar, M. F. B., Parma, J. O., and Miller, M. W. (2023). Does learning a skill with the expectation of teaching it impair the skill's execution under psychological pressure if the skill is learned with analogy instructions? Psychology of Sport and Exercise, 66, 102323. doi:10.1016/j.psychsport.2022.102323
Daou, M., Hutchison, Z. J., Bacelar, M. F. B., Rhoads, J. A., Lohse, K. R., and Miller, M. W. (2019). Learning a skill with the expectation of teaching it impairs the skill's execution under psychological pressure. Journal of Experimental Psychology: Applied, 25(2), 219-229. doi:10.1037/xap0000191
Do, H., Sonkar, S., and Sachan, M. (2026). Simulating students or sycophantic problem solving? On misconception faithfulness of LLM simulators. arXiv:2605.12748 [preprint].
Dunlosky, J., and Metcalfe, J. (2009). Metacognition. Thousand Oaks, CA: SAGE Publications.
Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., and Willingham, D. T. (2013). Improving students' learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4-58. doi:10.1177/1529100612453266
Fiorella, L., and Mayer, R. E. (2013). The relative benefits of learning by teaching and teaching expectancy. Contemporary Educational Psychology, 38(4), 281-288. doi:10.1016/j.cedpsych.2013.06.001
Fiorella, L., and Mayer, R. E. (2014). Role of expectations and explanations in learning by teaching. Contemporary Educational Psychology, 39(2), 75-85. doi:10.1016/j.cedpsych.2014.01.001
Fiorella, L., and Mayer, R. E. (2016). Eight ways to promote generative learning. Educational Psychology Review, 28(4), 717-741. doi:10.1007/s10648-015-9348-9
Guerrero, T. A., and Wiley, J. (2021). Expecting to teach affects learning during study of expository texts. Journal of Educational Psychology. doi:10.1037/edu0000657
Jacob, L., Lachner, A., and Scheiter, K. (2021). Does increasing social presence enhance the effectiveness of writing explanations? PLOS ONE, 16(4), e0250406. doi:10.1371/journal.pone.0250406
Jin, H., Lee, S., Shin, H., and Kim, J. (2024). Teach AI how to code: Using large language models as teachable agents for programming education. Proceedings of CHI 2024. doi:10.1145/3613904.3642349
Karpicke, J. D., and Blunt, J. R. (2011). Retrieval practice produces more learning than elaborative studying with concept mapping. Science, 331(6018), 772-775. doi:10.1126/science.1199327
Kobayashi, K. (2019). Interactivity: A potential determinant of learning by preparing to teach and teaching. Frontiers in Psychology, 9, 2755. doi:10.3389/fpsyg.2018.02755
Kobayashi, K. (2019). Learning by preparing-to-teach and teaching: A meta-analysis. Japanese Psychological Research, 61(3), 192-203. doi:10.1111/jpr.12221
Kobayashi, K. (2022). The retrieval practice hypothesis in research on learning by teaching: Current status and challenges. Frontiers in Psychology, 13, 842668. doi:10.3389/fpsyg.2022.842668
Kobayashi, K. (2024). Interactive learning effects of preparing to teach and teaching: A meta-analytic approach. Educational Psychology Review, 36(1), Article 26. doi:10.1007/s10648-024-09871-4
Koh, A. W. L., Lee, S. C., and Lim, S. W. H. (2018). The learning benefits of teaching: A retrieval practice hypothesis. Applied Cognitive Psychology, 32(3), 401-410. doi:10.1002/acp.3410
Lachner, A., Hoogerheide, V., van Gog, T., and Renkl, A. (2022). Learning-by-teaching without audience presence or interaction: When and why does it work? Educational Psychology Review, 34(2), 575-607. doi:10.1007/s10648-021-09643-4
Lachner, A., Jacob, L., and Hoogerheide, V. (2021). Learning by writing explanations: Is explaining to a fictitious student more effective than self-explaining? Learning and Instruction, 74, 101438. doi:10.1016/j.learninstruc.2020.101438
Lalley, J. P., and Miller, R. H. (2007). The learning pyramid: Does it point teachers in the right direction? Education, 128(1), 64-79.
Letrud, K., and Hernes, S. (2018). Excavating the origins of the learning pyramid myths. Cogent Education, 5(1), 1518638. doi:10.1080/2331186X.2018.1518638
Nestojko, J. F., Bui, D. C., Kornell, N., and Bjork, E. L. (2014). Expecting to teach enhances learning and organization of knowledge in free recall of text passages. Memory and Cognition, 42(7), 1038-1048. doi:10.3758/s13421-014-0416-z
Renkl, A. (1995). Learning for later teaching: An exploration of mediational links between teaching expectancy and learning results. Learning and Instruction, 5(1), 21-36. doi:10.1016/0959-4752(94)00015-H
Ribosa, J., and Duran, D. (2022). Do students learn what they teach when generating teaching materials for others? A meta-analysis through the lens of learning by teaching. Educational Research Review, 37, 100475. doi:10.1016/j.edurev.2022.100475
Rogers, K., Davis, M., Maharana, M., Etheredge, P., and Chernova, S. (2025). Playing dumb to get smart: Creating and evaluating an LLM-based teachable agent within university computer science classes. Proceedings of CHI 2025. doi:10.1145/3706598.3713644
Roscoe, R. D., and Chi, M. T. H. (2007). Understanding tutor learning: Knowledge-building and knowledge-telling in peer tutors' explanations and questions. Review of Educational Research, 77(4), 534-574. doi:10.3102/0034654307309920
Song, J., Guo, Z., and Lin, J. (2026). Simulating novice students using machine unlearning and relearning in large language models. Proceedings of AIED 2026, Lecture Notes in Computer Science. doi:10.1007/978-3-032-29744-0_42
Wang, Y., Lin, L., and Chen, O. (2021). The benefits of teaching on comprehension, motivation, and perceived difficulty: Empirical evidence of teaching expectancy and the interactivity of teaching. British Journal of Educational Psychology, 91(4), 1275-1290. doi:10.1111/bjep.12416
Wang, C., Petrie, C., Stouras, M., Ettlin, N., George, A., Mejia-Domenzain, P., Swamy, V., Kaeser, T., and Svensson, O. (2026). Turning 500+ students into teachers: A semester-long study of an AI teachable agent in an undergraduate algorithms course. Proceedings of the Thirteenth ACM Conference on Learning @ Scale, 350-360. doi:10.1145/3774398.3811623
Xu, Z., Zhang, J., Tang, A., and Lee, Y.-C. (2026). Who you explain to matters: Learning by explaining to conversational agents with different pedagogical roles. Proceedings of CHI 2026. doi:10.1145/3772318.3790298
