Writing for Others, Learning for Yourself: Teaching Expectancy, Explaining to an Absent Audience, Creating Teaching Materials, and Watching Dialogues

🌏 中文版

This is Part 22 of the series "Writing Articles That Teach," in its fifth section, "who you learn with, and who can't read you." The earlier posts in this section covered what a solitary reader misses (Part 20) and comment sections and communities. This one turns to the other end: the writer.

The trigger was this site's course guides. They take the caption tracks and slides of English-language courses and explain the material again, in Chinese, for people who never took the course. A common claim online is that teaching is the best way to learn. swyx's Learn In Public puts public blogging, cheat sheets, and answering forum questions first among all its advice, and asks the reader, "Did I mention that teaching is the best way to learn?" If that holds, writing a course guide is itself a way to learn. But writing an article is not teaching face to face. The reader is absent, asks nothing, and never looks confused. So the question here is: without live interaction, how much of the benefit of teaching others is left? There is a related question on the reader's side: can an article let readers learn the way people learn from watching someone else being taught?

Evidence on explaining things to yourself (self-explanation) is in Part 1; generative activities and the ICAP framework are in Part 8. This post covers only the extra piece: the audience is someone else.

How the sources were chosen, and their biases. I first scanned the literature, then picked sources to read closely. The rule was that every conclusion that uses a number needs a source I could read end to end, and at least one source with an opposite or null result. The final set: one meta-analysis (a study that pools many experiments) of "expecting to teach" and "actually teaching," two reanalyses of it by the same author, a review of mechanisms, and a commentary; a theoretical review of explaining to an absent audience; a meta-analysis of creating teaching materials for others; one teaching-expectancy experiment; three studies of watching someone else being taught; and one practitioner essay. The biases are obvious. First, five of the sources come from one researcher, Keiichi Kobayashi, and three from Michelene Chi's group, so "several studies agree" does not mean several independent teams found it. Second, almost all of it is college students in a lab, reading a few pages of science exposition, tested between tens of minutes and a week later. Third, in no study does "teaching" mean writing a public article for strangers. Every item in the references was read end to end, with an extraction record. Where I know a study exists but could not read a legal public full text, I say so and do not cite its results.

The short version

The g, d, and Ribosa's SMD in the table are all standardized mean differences (the Qian row is in percentage points); around 0.2 counts as small and around 0.5 as medium. High heterogeneity means the studies' results differ a lot from each other, so the average is only a summary.

What "teaching" means in these studies

Research in this area splits "teaching others" into pieces and tests each:

• Expecting to teach (preparing to teach): before reading, participants are told they will teach the material to another student afterward. After reading they do not actually teach; they take a test. The control group is told it will take a test. This isolates the intention to teach.
• Actually teaching: after reading, participants really teach. There are two forms. Teaching another student face to face counts as direct, or interactive, teaching. Writing an explanation, or recording an explanation on camera for an imagined student, counts as indirect, or non-interactive, teaching.
• Creating teaching materials: producing something others can learn from on their own, such as a video, a diagram, quiz questions, or a wiki page.

Writing a course guide is closest to indirect teaching and to creating materials. But keep one gap in mind from the start. Indirect teaching in these studies usually means a college student reads a short expository text in the lab, then writes an explanation or records one, for a fictional student who, participants are told, will watch or read it later, with limited preparation time and no chance to revise. Lachner and colleagues' review explicitly excluded studies of student-made websites or YouTube videos, because most had no control group, no element of teaching someone, or were made collaboratively. The meta-analysis of creating teaching materials included 23 studies; only 3 had a real online audience or real peers.

Only expecting to teach

The meta-analysis: small to medium effects in adults

Kobayashi's 2019 meta-analysis included 28 studies. Groups that only expected to teach, without teaching, compared with groups that expected a test, came out at g = 0.35 overall. Surface learning (recall, verbatim memory) was 0.39 and deep learning (inference, comprehension, far transfer) was 0.31; the author describes the two as roughly alike.

Two conditions belong with that number:

• It shows up only in adults. Samples aged 18 and over came out at 0.41. The 4 comparisons from elementary and secondary school came out at 0.04, with a confidence interval crossing zero: no effect detected. This appears only in the limitations section, not in the abstract.
• Expecting to teach face to face beat expecting to write or record: 0.50 versus 0.27. But the author notes that the number of studies is small and that the effects of test timing and teaching type may be confounded with other methodological features of the included studies.

The control groups are not consistent either. Most expected a test, but some expected to write a summary, list keywords, or solve problems. Ten of the 28 comparisons had effects near zero or negative.

What one experiment looks like

Nestojko and colleagues' two experiments from 2014 are typical of this work. UCLA undergraduates read a passage of about 1,500 English words for 10 minutes, with no note-taking. Half were told they would teach the content to another participant, who would then take a test; the other half were told they would take a test. After reading, everyone did an unrelated task for 25 minutes. Then the teaching group was told that the student they were supposed to teach had failed to show up, and took the same test as the others. Nobody actually taught.

• Experiment 1 (56 people): the expect-to-teach group wrote down more in free recall (d = 0.56), recalled it in an order closer to the passage's paragraph structure (d = 0.65), remembered more main points, and did better on a short-answer test. On a second coherence measure that does not depend on the passage's structure, no difference was detected.
• Experiment 2 (44 people, using a textbook section): no difference was detected in overall fill-in-the-blank scores. Only on main-point questions did the expect-to-teach group trend toward doing better (p = .083). Under a different rule for classifying the questions, the expectancy-by-question-type interaction was no longer significant, though the pattern in the numbers stayed the same.

The authors' explanation is that people who expect to teach read the way a teacher prepares a lesson: picking out main points and organizing the structure. That is an inference; the study did not measure the reading process. The authors close with:

Students seem to have a toolbox of effective study strategies that, unless prodded to do so, they do not use.

They also acknowledge that this setup, in which participants expect to teach but never do, cannot move into a classroom, because students would soon notice they never actually have to teach.

No effect detected in Japanese students or under-18s

In another 2019 report, Kobayashi split the meta-analysis data by sample. For the 6 comparisons with Japanese students, the effect of expecting to teach was g = 0.09, with a confidence interval crossing zero; other samples came out at 0.40, and the difference between them was significant.

This result has an obvious confound. Five of the 6 Japanese comparisons involved expecting to teach indirectly (writing or recording), and indirect expectancy has a smaller effect to begin with. The author dealt with this and concluded that the confound explains only part of it: after excluding Japanese students, the effect of indirect expectancy was still above zero and clearly higher than the Japanese figure. Both sides need saying: the confound is real, and the author judges it insufficient to explain the null.

The same report has an important contrast. When participants expected to teach and then actually taught, Japanese students came out at g = 0.58 and other samples at 0.51, with no meaningful difference. So it would be wrong to say that Japanese students don't learn from teaching others; only the "expecting alone" piece is unusual. This is a university bulletin, not a peer-reviewed journal; the evidence comes from 6 comparisons by a handful of researchers, and the author calls for replication by others.

Lachner and colleagues' review is more cautious about teaching expectancy: results are mixed, and the moderation effect Fiorella and Mayer reported in 2014 (recording an explanation afterward worked only with a teaching expectancy) often failed to replicate in other studies. The authors suspect this is because the expectancy effect is small and the samples too small to detect it.

There is also a newer 2024 meta-analysis by Kobayashi (Educational Psychology Review) that directly compares teaching with and without a prior expectancy. The study exists, but I could not read a legal public full text, so I do not cite its numbers.

Writing or recording for an absent audience

Indirect teaching in the meta-analysis

In the same meta-analysis, expecting to teach and then actually teaching came out at g = 0.56 overall. That number is easy to take as "the effect of teaching others," but in 12 of the 16 comparisons the "teaching" was writing an explanation or recording a video, with no live person present. Split apart:

• Expecting to teach, then writing or recording: g = 0.48 (12 comparisons)
• Expecting to teach, then teaching face to face: g = 0.84 (4 comparisons)

In a 2019 perspective article, Kobayashi reanalyzed the same data and added one row: no teaching expectancy, asked to write or record only after reading, g = 0.23, from only 2 comparisons. The three rows in descending order are face to face 0.84, indirect with expectancy 0.48, and indirect without expectancy 0.23, and the differences are significant.

Two things to keep in mind with these three numbers. First, in all three rows the control group neither expected to teach nor taught; the rows are not compared with each other directly. The author says that, to their knowledge, only two experiments directly compared face-to-face teaching with recorded teaching. Second, this is a reanalysis inside a perspective article; 0.84 rests on 4 comparisons and 0.23 on 2. The author's conclusion in the same article:

to date, there is no single satisfactory explanation for why interactivity affects learning by preparing to teach and teaching

A study-by-study review

Andreas Lachner, Vincent Hoogerheide, Tamara van Gog, and Alexander Renkl's 2022 review is specifically about explaining to an absent audience, defined as explaining, after studying, to a real but absent, remote, or imaginary audience, without interaction. The authors deliberately did not run a meta-analysis and pooled no effect sizes; the aim was to lay out mechanisms and boundary conditions. So it cannot tell you how large the effect is.

Its value is Table 1, which lists comparisons and results for 27 rows, study by study. I sorted the rows myself (the grouping is mine, not a statistic in the paper): in about 12 rows the explaining group did better on at least one outcome, in 9 rows no difference was detected on any outcome, and in 6 rows the explaining group did worse than the comparison on at least one outcome; a few rows had both positive and negative results. The worse rows are worth noting:

• Roscoe and Chi 2008: the group that explained to a camera did worse on comprehension than the group that explained to themselves (d = 1.64).
• Experiment 2 of Lachner and colleagues 2021 (pre-service teachers): self-explanation beat explanations written for a fictional audience (d = 0.59).
• In two unpublished studies, recorded explanations lost to simply restudying.

Two other findings from the review bear on writing guides:

• Type of material. Studies with procedural material, such as troubleshooting, step-by-step procedures, and motor skills, almost all detected no effect, and the one exception did not replicate. The authors therefore recommend using this approach when learning complex concepts from expository text.
• More than retrieval. Explaining is often described as just a form of retrieval practice (pulling what you learned out of memory). The authors conclude that the current evidence suggests it is more than retrieval. Yet in the same table, the explaining groups in Lachner 2021 and Jacob 2021 did not beat groups that only practiced retrieval.

The authors list many caveats themselves: most studies measured only immediate effects; a lot of unreviewed gray literature was included (Table 1 has 7 unpublished rows, while the methods section gives peer-reviewed publication as an inclusion criterion, and the two do not match); and every study used a "stripped down" version of explaining, with limited preparation time, no notes, and no chance to revise or re-record.

Speaking or writing

The review's model says there is considerable evidence for speaking. Study by study, the direct comparisons of speaking and writing are: Hoogerheide and colleagues 2016 found no direct difference between the two, only that speaking beat restudying while writing did not; Lachner and colleagues 2018 found speaking better on transfer (d = 0.67), with no difference detected on recall; Jacob and colleagues 2020 found speaking better only for the difficult text, on transfer (d = 0.32); Jacob and colleagues 2021 and another Jacob study detected no difference on recall or transfer. The more accurate summary: a few studies found speaking better on transfer questions, no difference on recall, and some found no difference at all. Between audio and video recordings, the two studies that compared them directly detected no difference.

When to explain

Lachner and colleagues 2020 compared explaining after reading a whole section with explaining halfway through. Overall, explaining did not beat retrieval, but the group that explained midway remembered more than the group that explained at the end (η²p = .06); no difference was detected on transfer. The authors' analysis indicates the difference ran through monitoring whether one had understood.

Does the audience need to be there

Intuitively, an explainer who feels someone is listening should do better. The studies in the review point the other way: explaining to a real person on the other end of a video call led to worse immediate and delayed problem-solving scores than just studying the worked example, while explaining to a fictional audience neither helped nor hurt. In the authors' words:

social presence during non-interactive teaching may be regarded as a double-edged sword, as high levels of social presence impaired rather than contributed to learning.

The authors also caution that none of these experiments measured how much social presence participants actually felt; there was no manipulation check, so the results need careful interpretation. Writing an explanation in a chat room (with added social cues) and writing it in a text editor produced no detected difference in test scores.

Creating teaching materials for others

Jesús Ribosa and David Duran's 2022 meta-analysis included 23 studies and 62 comparisons and asked how much students learn about the content when they create teaching materials for others. The control groups had business-as-usual instruction or did another activity that produced no materials (restudying, retrieval practice, face-to-face tutoring, and so on). Under a random-effects model, the average effect was 0.17 (95% CI 0.04–0.31).

Some moderator results:

• What the control group did matters a lot. When the control group did an activity known to be effective (such as retrieval practice), the effect was about 0.31 smaller than when the control group only restudied. The abstract reads as if the effect was not significant against effective activities, but the paper reports only the difference between the two kinds of control, not whether each was significant on its own.
• Making audiovisual or visual materials beat making text, with estimated differences of 0.63 and 0.71. The authors themselves caution that there were only 7 studies of text and the results should be read carefully: in one, writing multiple-choice questions lost to restudying; a small wiki study detected no difference; writing explanations did no better than restudying or retrieval.
• Access to the source material. Creating materials with full access to the source material had a smaller effect than with no access (difference 0.34, p = .032); the other two contrasts (full versus limited, limited versus none) detected no difference. This is a cross-study comparison, not an experimental manipulation. I noticed that the "full access" group contains several of the most negative comparisons, and those studies mostly had effective activities as controls; the authors say the difference stayed significant after controlling for control type, but did not rule out other confounds one by one.
• Educational level and subject showed no detected difference.
• Publication bias. The abstract says none was detected; the body says 24 of the 62 comparisons fall outside the funnel plot's limits, and only the Egger test (p = .095) fell short of .05.

Kobayashi's 2023 commentary (an opinion piece, no new data) raised two problems with this meta-analysis. First, several included studies never told students beforehand that their work was for others, so a strict reading of the inclusion criteria would have excluded them. Second, the 0.17 mixes four types of comparison: in some, the control group was also told to teach and simply made no materials. Studying with the intention to teach is already partly creating materials, so comparing the full process of creating materials against a partial one underestimates the effect. The author agrees that creating materials can lead to effective learning, but argues that the scope of "creating teaching materials" is not clearly defined, and proposes defining it by the intention to teach.

My own reading (not a limitation the authors list): multiple comparisons from the same study were treated as independent studies, and they share control groups, which makes the standard errors too small.

Why it might work: five hypotheses

Kobayashi's 2020 review (in Japanese; a narrative review with no stated search method) organizes the explanations researchers have offered for why teaching others works into five hypotheses and checks each against existing studies:

The author says that even the two supported hypotheses lack decisive, sufficient evidence; most studies show only that teaching relates to some process and that the process relates to scores, without testing what lies between.

Two details from this review bear directly on writing:

• Most of what explainers produce is not knowledge building. The author compiled the share of "knowledge-building explanations" among all explanations in several studies: 7% to 39% for face-to-face teaching and 13% to 32% for recorded teaching. Most of the rest restates what was read. The studies count in different units, so only the rough picture holds.
• Monitoring is triggered more easily by interaction. After a student asks a deeper question, the explainer often follows with statements that monitor their own understanding. In Roscoe and Chi 2008, about 45% of face-to-face explanations and self-explanations contained monitoring, versus 25% of recorded ones. The author therefore speculates that if monitoring requires interaction, the metacognition hypothesis applies mainly to face-to-face teaching and much less to recording or writing explanations. This is the author's speculation.

On the reader's side: watching someone else being taught

So far the question has been how much the writer learns. The reader's side has a parallel question: if an article is written as "a student being taught," can readers learn the way people learn from watching someone else being taught? Part 20 also uses these studies; here I cover only what bears on the form of an article.

In the lab: watching in pairs and solving together. Chi, Roy, and Hausmann's 2008 study had 70 University of Pittsburgh undergraduates learn to solve Newton's-laws problems, in five groups. On a posttest about a week later, gains in deep problem-solving steps were: tutored one on one 21.3%; watching a tutoring video in pairs while solving the same problems 17.0%; solving in pairs without the video 7.2%; watching the video alone 8.9% (no gain detected); studying alone 3.4% (no gain detected). No difference was detected between the tutored and the pair-watching groups.

This result has to be read with its conditions. The pair-watchers were also solving the same problems on a whiteboard, and could discuss, pause, and rewind; they were not just watching. Which student they watched also mattered: watching a strong student's session gave a 21.9% gain, a weak student's 7.8%. In the watch-alone group, the authors identified 4 more active participants: they wrote more problem-solving steps on the whiteboard (74 versus 7.3 on average), paused and rewound the tape more often, and questioned themselves more often (for example, asking whether a step was allowed). Those 4 gained 18.4% and the other 6 gained 2.6%, with similar pretest scores (a post hoc split with no statistical test). The tutored group was not randomly assigned, and there was only one tutor. The more accurate statement: people who watched alone and passively learned almost nothing; people who watched and worked together showed gains with no detected difference from the tutees. That does not mean watching is as good as being tutored, because the study ran no equivalence test, and the tutored group's numbers were still higher.

Dialogue videos versus monologue videos. Chi, Kang, and Yaghmourian's 2017 paper reanalyzes data from another study (Muldner and colleagues 2014). The same set of tutors each recorded two videos: a dialogue version tutoring a student, and a lecture-style monologue to the camera. College students watched in pairs and worked through a workbook together. On 11 transfer questions, the dialogue-watching pairs learned more than the monologue-watching pairs (t(38) = 2.05, p < .05). The effect size printed in the paper does not match this test statistic, so I do not cite it. The dialogue watchers solved more workbook problems themselves (16.7 versus 10.9 out of 22), and about half of the monologue watchers' answers were copied from the video. People who watched the dialogue videos also discussed what the student in the video said more often than what the tutor said. After segments in which the student said something wrong and the tutor corrected it, the watchers discussed more.

A real online course. Qian, Hong, and Chi's 2024 study moved the same design into online courses at a US public university, with the videos part of the course and the posttest a real exam.

• Biology (416 people): posttest 69.07% after dialogue videos versus 65.65% after monologue videos, with the pretest as a covariate; the difference was significant, η² = 0.03. Both groups scored about 61% on the pretest, so the actual gap is about 3.4 percentage points.
• Calculus (74 people, two weeks): dialogue 57.80%, monologue 46.16%, η² = 0.11. But both groups scored only about 10% on the pretest, and 17% of participants were dropped for missing a test.

In both courses the dialogue videos were longer than the monologues (36 versus 26 minutes, 22 versus 15 minutes), and time was not controlled. Every condition filled in the same demanding worksheet, so there is no watch only, no worksheet comparison. In biology, watching alone and watching in pairs were two classes in different semesters, so that contrast cannot show that watching together is better; and the implicit-question scores behind that contrast do not add up to the totals in the paper's own score table.

All three studies tested video. Whether writing an explanation as a text dialogue (a student asks, gets stuck, gets corrected) helps readers learn more has not been tested directly by any study.

The claim that teaching is the best way to learn

swyx's essay is career advice, with no data and no cited research. It says three things about learning itself: the title calls it the fastest way to learn; you write for yourself three months ago, and the biggest beneficiary is your future self; and teaching is the best way to learn. Most of the rest is about mentors, reputation, job opportunities, and income. The output it encourages is also varied: blog posts, talks, Q&A, videos, open-source pull requests, cloning other people's projects from scratch.

Matching the essay against the research above is my own mapping; the essay does not draw these lines:

• Writing for the you of three months ago (the essay says "talk to yourself from 3 months ago") is closest to indirect teaching with a teaching expectancy, aimed at an imagined reader. In the research, this category (participants mostly college students) averages g = 0.48, but individual studies are unstable, and almost no effect was detected for step-by-step procedural content.
• Letting the internet correct you is the only interactive part. The nearest results are face-to-face teaching at 0.84 (4 comparisons) and the finding that monitoring is mostly triggered by students' questions. But comments are delayed and written, and research has not tested them. In the perspective article, Kobayashi lists asynchronous text exchanges among the forms of interaction not yet studied.
• No study has compared learning speed, so nothing supports "fastest."

Where the sources disagree

The following results point in different directions, or look contradictory while measuring different things. I lay them side by side without picking one.

Does teaching expectancy itself have an effect? Kobayashi's meta-analysis: g = 0.41 in adults. Kobayashi's reanalysis of Japanese students: 0.09, no effect detected. Lachner and colleagues' review: mixed results, and Fiorella and Mayer's moderation effect often failed to replicate. Nestojko and colleagues: an effect in Experiment 1, no overall difference detected in Experiment 2.

Is having someone present good or bad? Kobayashi's reanalysis: face-to-face teaching at 0.84 exceeds indirect teaching at 0.48, and the mechanisms review says monitoring is mainly triggered by interaction. The experiments Lachner and colleagues compiled: explaining to a real person over video was worse than just studying the worked example, and explaining to a fictional audience made no difference. The two sides measure different kinds of presence: the first is face-to-face teaching with back-and-forth, the second is explaining while a real person watches without interacting.

Do written explanations help? Kobayashi's meta-analysis lumps writing explanations and recording videos together, averaging 0.48. In Ribosa and Duran's meta-analysis, making written materials had a clearly smaller effect than making audiovisual or visual ones, and most of the 7 text studies had weak results. Lachner and colleagues' direct comparisons: a few studies found speaking better than writing, most found no difference.

Explaining while looking at the source, or from memory? Ribosa and Duran: full access to the source while creating materials had a smaller effect than no access. Lachner and colleagues recommend explaining from memory in their conclusion; but the same review notes that Hoogerheide 2019 found an effect even with the material available, the review contains no study that compares the two directly, and the authors list it as future research. Kobayashi's mechanisms review also cites studies in which recording explanations while looking at the material or notes still worked.

Is explaining just retrieval? Lachner and colleagues: current evidence suggests explaining is more than retrieval. In the same table, the explaining groups in Lachner 2021 and Jacob 2021 did not beat retrieval-only groups, and in Koh 2018, "explaining plus retrieval" and "retrieval only" did not differ.

Do people who watch someone being taught learn as well as the tutees? The abstract of Chi 2008 says the pair-watchers learned "just as effectively" as the tutees, and the abstract of Chi 2017 says they "can learn as well as the tutees." Both rest on no detected difference in small samples (10 versus 20 people), with no equivalence test, and the tutored group's numbers were still higher.

What you can do

Each item below is my (this site's) judgment, with the supporting results in parentheses. They are split between the writer and the reader.

If you want to learn by writing articles:

1. Decide who you are writing for before you start reading, not after you finish. (Basis: teaching expectancy in adults, g = 0.41; writing without an expectancy, 0.23 from only 2 comparisons. Counter-evidence: no effect detected in Japanese students or under-18s, and Lachner and colleagues say the expectancy effect is small and often fails to replicate.)
2. Stop halfway and explain once, instead of writing only after finishing. (Basis: in Lachner and colleagues 2020, the group that explained midway remembered more. One set of studies, with no difference detected on transfer.)
3. Close the source, write a passage, then open the source and check. (Basis: Ribosa and Duran's cross-study comparison and Lachner and colleagues' recommendation. Neither has a direct experimental comparison, and accuracy requires checking against the source, so I split the two into consecutive steps. Also, the illusion of explanatory depth covered in Part 15 shows that for mechanistic knowledge, writing out a step-by-step explanation makes self-ratings drop clearly. That is another reason to write first and check second, though it concerns self-ratings, not tests.)
4. If you can find a live listener, explain to them and let them ask questions. (Basis: face to face 0.84, 4 comparisons; monitoring is mostly triggered by the other person's questions. Counter-evidence: having a real person merely watch over video, without interaction, was worse than just studying.)
5. Don't expect to learn much from writing step-by-step procedural articles (installation, commands, debugging workflows). (Basis: the procedural-material studies compiled by Lachner and colleagues almost all detected no effect.)

If you design articles for readers:

6. Phrase prompts after a section as "explain in your own words why this step holds," not "pretend to teach this to someone who doesn't get it." (Basis: self-explanation has meta-analytic support for learning outcomes, see Part 1; its effect on judging whether you understood is unsettled, see Part 15. In Roscoe and Chi 2008 and Experiment 2 of Lachner 2021, self-explanation beat explaining to an absent audience, though Lachner and colleagues note that Roscoe and Chi's comparison is confounded by when the explaining happened. How much a different-person audience adds is unstable across studies.)
7. If you want to write an explanation as a dialogue, treat it as an experiment, not an evidence-backed practice. (Basis: dialogue videos beat monologue videos in the lab and in one real course, but text dialogues have not been tested directly; and in the studies, the watchers were also working through a workbook.)

Checking this site's practice

This site's course-guide spec is .agents/skills/post/references/course-guide-contract.md. Four points relate to this post:

One more point about how this site actually works. Many of the course guides are first drafted by AI agents from slides and captions; Part 29 will cover that process. All the studies above measure how much the person who does the explaining learns. If an agent does the explaining, the part of the benefit that comes from teaching others at least does not automatically go to the site's author; whether reviewing, editing, and checking a draft counts as teaching has not been studied. This is my inference.

Three things to do tonight

1. Pick a topic you are learning this week, and before reading, write one line: "After reading, I will explain what, to whom."
2. Stop halfway, close the source, and spend five minutes writing an explanation for that person; then reopen the source and mark what you got wrong or left out.
3. Go back to a teaching article you have written, and replace a closing "try teaching this to someone" with a question of the form "explain in your own words why…", plus a criterion for a good answer.

Limits of this evidence

• No study has tested writing a public article for strangers. The closest is writing an explanation or recording one for a fictional classmate in the lab. Extending the results to a guide several thousand words long, written over several days and revised repeatedly, is my inference. Lachner and colleagues themselves say every study used a "stripped down" version of explaining.
• Materials are mostly short science exposition, participants mostly college students, tests mostly within a week. In Kobayashi's meta-analysis, all delayed tests were within a week; most studies Lachner and colleagues compiled had only immediate tests.
• The researchers are concentrated. Kobayashi's meta-analysis, the two reanalyses, the mechanisms review, and the commentary all come from one author; the three studies of watching someone being taught come from one research group.
• Heterogeneity is high. I² in Kobayashi's meta-analysis ranges from 76% to 94%, and the author says the sources of heterogeneity remain largely unclear.
• Publication bias is only partly addressed. Kobayashi's meta-analysis uses only the fail-safe N (Learning How to Learn explains why that measure overstates robustness). In Ribosa and Duran's funnel plot, 24 comparisons fall outside the limits, and the Egger test did not reach significance.
• My own observation: several confidence intervals in Kobayashi's meta-analysis and perspective article are unusually narrow given such high heterogeneity (for example, 0.56 with I² = 93% has an interval of only [0.52, 0.61]); the methods section says only that results were weighted by sample size. I can't tell from the paper how they were computed, so in the main text I use the intervals only to judge whether they cross zero (if the true intervals are wider, the ones that cross zero still do). The subgroup difference tests (Japanese students versus others, face to face versus indirect, 0.50 versus 0.27) come from the same variance calculation and may also lean toward significance.
• Secondhand results: the results of Roscoe and Chi 2008; Lachner and colleagues 2018, 2020, and 2021; Jacob and colleagues 2020 and 2021; Hoogerheide and colleagues 2016; Koh and colleagues 2018; Fiorella and Mayer 2014; and Muldner and colleagues 2014 were all read through reviews, meta-analyses, or reanalyses. I did not read the original studies.
• Known to exist, but no legal public full text read: Kobayashi's 2024 meta-analysis. I do not cite its numbers or list it in the references.
• My inferences: mapping indirect teaching in the studies to writing course guides; mapping swyx's claims to teaching expectancy and interaction; the seven items under "What you can do" and the three things to do tonight; the paragraph about AI-drafted guides. Each is marked in the text.

Appendix: study designs and numbers

Kobayashi (2019), meta-analysis, Japanese Psychological Research 61(3). 28 studies; expecting to teach: 28 comparisons, 1,283 people; expecting and actually teaching: 16 comparisons, 751 people. 24 studies at age 18 and over, 4 in elementary and secondary school. Random-effects model, Hedges' g. Expecting to teach g = 0.35 (95% CI 0.27–0.44), I² = 76%; expecting and actually teaching 0.56 (0.52–0.61), I² = 93%. Expecting to teach: surface 0.39 (k = 17), deep 0.31 (k = 20), immediate 0.28 (k = 20), delayed 0.53 (k = 12). Expecting and actually teaching: surface 0.58, deep 0.55, immediate 0.51, delayed 0.60. Expecting face to face 0.50 (k = 10) versus expecting to write or record 0.27 (k = 18), QB(1) = 6.51, p < .05; actually face to face 0.84 (k = 4) versus writing or recording 0.48 (k = 12), QB(1) = 47.95, p < .001. Expecting to teach under 18: 0.04 (−0.16 to 0.25, k = 4); 18 and over 0.41 (0.32–0.50, k = 24). Fail-safe N 667 and 285. One sentence in the text says the immediate effect was larger than the delayed effect, contradicting the numbers in the same paragraph and in Table 3, where delayed is larger; I follow the table. In Table 2, the footnotes for the two Ito and Kakihana 2009 rows do not match the teaching-type column, and the paper has no correction.

Kobayashi (2019), perspective article, Frontiers in Psychology 9. A reanalysis of the data above, adding two comparisons with no expectancy in which participants wrote or recorded afterward: Fiorella and Mayer 2014 Experiment 2 (g = 0.21) and Hoogerheide and colleagues 2016 Experiment 1 (g = 0.24). Table 2: expecting indirect 0.27 (k = 18), expecting direct 0.50 (k = 10); indirect without expectancy 0.23 (k = 2), indirect with expectancy 0.48 (k = 12), direct with expectancy 0.84 (k = 4), three-way difference QB(2) = 118.79, p < .001. The two Qw values in Table 2 are swapped relative to Table 4 of the 2018 meta-analysis, so one of the two is mislabeled. In the same table, the Qw values for the 0.48 and 0.84 rows also differ from the meta-analysis (117.22 versus 114.98, 44.29 versus 43.41), and the paper does not explain why. Open questions the author lists: no study has used a crossed design (expecting indirect teaching but teaching directly); no study has examined asynchronous text, phone, or video-call interaction; the argument rests mainly on artificial experimental settings.

Kobayashi (2019), reanalysis of Japanese students, Bulletin of the Faculty of Education, Shizuoka University 70. Adds 4 newer studies to those in the 2019 meta-analysis, 31 comparisons in all (6 Japanese, 25 other). Overall g = 0.35 (0.27–0.43), I² = 74%; Japanese students 0.09 (−0.11 to 0.29), others 0.40 (0.31–0.49), QB = 7.63, p < .01. Excluding Japanese students: expecting direct teaching 0.57 (0.40–0.73), expecting indirect teaching 0.31 (0.16–0.45), QB = 4.90, p < .05. Expecting and actually teaching: Japanese students 0.58 (k = 4), others 0.51 (k = 13), QB < 1. In 1 of the 6 Japanese comparisons the control group generated keywords rather than expecting a test. The labels Fukaya 2013a and 2013b are swapped between this report and the 2019 meta-analysis.

Kobayashi (2020), mechanisms review, Japanese Journal of Educational Psychology 68(4), in Japanese. Narrative review, with no search method or inclusion criteria, excluding collaborative-learning studies. Five hypotheses: knowledge building; motivation (three sub-hypotheses: intrinsic motivation, the protégé effect, social presence); generating explanations; metacognition; retrieval practice. Shares of knowledge-building explanations in Table 1: Ito and Kakihana 2009 face to face 37%, recorded 32%; Ito and Kakihana 2019 face to face 39%; Lachner 2020 recorded 14% and 19%; Roscoe 2014 face to face 7%; Roscoe and Chi 2008 face to face 26%, recorded 13%. The studies use different denominators (statements, idea units, paragraphs).

Kobayashi (2023), commentary, Frontiers in Psychology 14. Opinion piece. Splits the comparisons in Ribosa and Duran into four 2×2 types (whether the intervention group expected to teach × whether the control group did), gives one example of each, and points out that in two types the control group was also preparing to teach, which underestimates the effect. Proposes defining material creation by the intention to teach, and including preparing to teach in the definition. The author says there are currently no data on preparing while creating.

Lachner, Hoogerheide, van Gog, Renkl (2022), Educational Psychology Review 34. Theoretical review, no meta-analysis. Searched Google Scholar and Web of Science, getting 236 hits in March 2020, plus scholars' publication lists, gray literature, and additions during review; the text says 25 articles, while Table 1 has 27 rows (20 published, 7 unpublished). Excludes teachable-agent studies and studies of student-made websites, videos, and similar products. Selected Table 1 results: Fiorella and Mayer 2013 immediate test, teaching expectancy d = 0.59, recorded explanation d = 0.82; Hoogerheide and colleagues 2014 Experiment 2, recorded explanation versus test expectancy, recall d = 1.10; Lachner and colleagues 2018 speaking versus writing, transfer d = 0.67; Jacob and colleagues 2020 difficult text, explaining versus retrieval d = 0.36, speaking versus writing d = 0.32; Roscoe and Chi 2008 self-explanation versus recorded explanation d = 1.64; Lachner and colleagues 2021 Experiment 2, self-explanation versus written explanation d = 0.59; Hoogerheide, Renkl and colleagues 2020 (unpublished) Experiment 1, restudying versus explaining to a real person over video, transfer d = 0.70; Wassenburg and colleagues 2020 (unpublished), restudying versus recorded explanation, comprehension d = 0.62. The model's outcome box lists only conceptual learning. The four recommendations in the conclusion: explain aloud, explain from memory, study with a teaching expectancy, and restudy after explaining. Of these, "from memory" has no direct comparison anywhere in the review, and "with a teaching expectancy" is described in the body as having mixed results.

Ribosa and Duran (2022), Educational Research Review 37. Searched PsycInfo, Scopus, and Web of Science in October 2021, without "learning by teaching" as a search term; the databases yielded only 3 eligible studies, and the rest came from the authors' scoping review (7), Lachner and colleagues' 2022 list (12), and a scholar's recommendation (1), for 23 studies and 62 comparisons; after removing duplicate groups, 1,723 people in creation groups and 1,774 in control groups. Peer-reviewed journals only. Random effects 0.17 (p = .013, 95% CI 0.04–0.31, the CI shown only in the forest plot); common effect 0.22 (0.16–0.27); Q = 296.80, I² = 79.4%. Product: audiovisual versus text 0.627 (SE 0.141), visual versus text 0.705 (SE 0.233), audiovisual versus visual −0.079 (no difference). Source access: full versus none −0.337 (p = .032), full versus limited −0.317 (p = .109), limited versus none −0.020 (p = .915). Control: beneficial versus non-beneficial −0.305 (p = .027). Year of publication, research design, educational level, and subject showed no detected difference. Limitations the authors list: gray literature excluded, heterogeneous control groups, tests closely aligned with the intervention, varying test timing, and not all tests checked for reliability.

Nestojko, Bui, Kornell, Bjork (2014), Memory & Cognition 42(7). Experiment 1: 56 UCLA undergraduates, a 1,541-word passage, 10 minutes of reading, a 25-minute distractor task. Proportion of correct idea units in free recall .17 versus .13 (t(54) = 2.08, p = .043, d = 0.56); recall efficiency 1.50 versus 1.02 units per minute (d = 0.79); match between recall order and passage structure .81 versus .67 (d = 0.65); between-sentence coherence .24 versus .22 (p = .380); main points .33 versus .24 (d = 0.56); no difference detected for important or unimportant details; short-answer test, main effect of expectancy F(1, 54) = 5.04, p = .03. Experiment 2: 44 people, a 1,300-word textbook section, 9 minutes of reading, a 30-minute distractor task, background knowledge as a covariate. Fill-in-the-blank .44 versus .39, F(1, 41) = 0.76, p = .389; the expectancy-by-question-type interaction was significant (p = .009), and not significant under a 66% rater-agreement classification; main-point questions p = .083, detail questions p = .814. No delayed test.

Chi, Roy, Hausmann (2008), Cognitive Science 32(2). 70 University of Pittsburgh undergraduates who had taken physics in high school but not in college. Five groups: tutored 10, watching together 20 (10 pairs), solving together 20, watching alone 10, studying alone 10; the first 10 recruited were all assigned to tutoring. One tutor, a physics PhD who had taught college physics for over 30 years. Intervention capped at 2 hours; posttest an average of 7.4 days later. Gains on matched deep steps: 21.3%, 17.03%, 7.19%, 8.91% (not significant), 3.39% (not significant); five-group difference F(4, 65) = 3.787, p = .008. Watching together versus alone p = .055 (d = .425). Strong-student tape 21.9% versus weak-student tape 7.8%; the 6 evenly interacting pairs 26% versus the 4 uneven pairs 4%. The 10 watching pairs were analyzed as n = 20 individuals, though the two members of a pair are not independent, and the paper does not handle this; several F values and p values do not correspond.

Chi, Kang, Yaghmourian (2017), Journal of the Learning Sciences 26(1). A reanalysis of Muldner and colleagues 2014: college students, 10 tutees and 40 watchers (same-gender pairs, 10 pairs per condition), the concept of diffusion, 5 tutors (only 1 experienced), videos averaging 21 minutes (monologue) and 25 minutes (dialogue), a difference that was not significant. Watchers co-wrote one workbook per pair, with no time limit and free control of playback. On 11 transfer questions: tutees' pre-to-post gain d = 1.07, dialogue watchers' d = 1.08; both are within-group gains, not between-group comparisons; monologue watchers' gain p = .26, d = 0.36. Dialogue versus monologue watchers t(38) = 2.05, p < .05; the effect size printed in the paper does not match the t value, so I do not cite it. Between-group comparisons on the 18 similar questions are not reported. Workbook problems solved independently: dialogue 16.70, monologue 10.90 (out of 22); solved jointly 10.50 versus 5.30. Correlation between each watcher's number of substantive comments and gain r = .50 (p = .03, 20 pairs); the other behavior correlations were marginal or not significant. The abstract says behavior "mediated" learning; the body twice says the sample was too small and no mediation analysis was run.

Qian, Hong, Chi (2024), International Journal of STEM Education 11. Stratified random assignment; 6 pairs of videos per course, same instructor, same slides, with the dialogue version tutoring a student who had taken the course the previous semester. Study 1 (biology): 416 people, watching individually in the first semester (114 per condition) and in pairs in the second (94 per condition). Posttest: monologue 65.65, dialogue 69.07, F(1, 411) = 11.21, η² = 0.03; watching individually η² = 0.03, watching in pairs η² = 0.02. Study 2 (calculus): 74 people analyzed, pretest 10.97% versus 8.55%, posttest 46.16% versus 57.80%, η² = 0.11; implicit questions 4.42 versus 5.76 correct, η² = 0.10; no difference detected on explicit questions. The paper has no limitations section. Converting the biology individual-watching explicit plus implicit question counts gives about 56.6%, which does not match the 66.93% in Table 2. Pairs were analyzed as individuals.

swyx, Learn In Public, personal website, page dated 2018-06-19, updated since (it mentions 2022). No data, no cited research.

References

Every item was read end to end, with an extraction record; the version read is given in parentheses.

• Kobayashi (2019), Learning by Preparing-to-Teach and Teaching: A Meta-Analysis (published version, free to read on Wiley's site).
• Kobayashi (2019), Interactivity: A Potential Determinant of Learning by Preparing to Teach and Teaching (open-access published version).
• Kobayashi (2019), 日本の学生には教授予期の学習効果が見られない？ (Teaching Expectancy May Not Foster Japanese Students' Learning) (bulletin PDF in Shizuoka University's institutional repository; in Japanese).
• Kobayashi (2020), 他の学習者に教えることによる学習はなぜ効果的なのか？——5 つの仮説とそれらの批判的検討—— (Why is learning by teaching other students effective? An examination of five hypotheses) (published version on J-STAGE, in Japanese; doi:10.5926/jjep.68.401).
• Kobayashi (2023), Learning by creating teaching materials: Conceptual problems and potential solutions (open-access published version).
• Lachner, Hoogerheide, van Gog & Renkl (2022), Learning-by-Teaching Without Audience Presence or Interaction: When and Why Does it Work? (open-access published version; references only skimmed).
• Ribosa & Duran (2022), Do students learn what they teach when generating teaching materials for others? A meta-analysis through the lens of learning by teaching (open-access publisher page; tables and forest plot checked against the published version appended to the first author's doctoral thesis in the Autonomous University of Barcelona repository; if the repository is unreachable, see the Wayback archive of the thesis PDF).
• Nestojko, Bui, Kornell & Bjork (2014), Expecting to teach enhances learning and organization of knowledge in free recall of text passages (published version readable without login on Springer's site; figure values not read).
• Chi, Roy & Hausmann (2008), Observing Tutorial Dialogues Collaboratively: Insights About Human Tutoring Effectiveness From Vicarious Learning (published version, free to read on Wiley's site, without an open license; online appendices not read).
• Chi, Kang & Yaghmourian (2017), Why Students Learn More From Dialogue- Than Monologue-Videos: Analyses of Peer Interactions (read as the published-version PDF on the first author's lab website at Arizona State University).
• Qian, Hong & Chi (2024), Learning from watching dialog and monolog videos in online STEM courses (open-access published version; figure values not read).
• swyx, Learn In Public: The fastest way to learn (current version on the author's site).
• Posts on this site: Writing the Examples and Openings of a Teaching Article, Writing Questions and Wording in a Teaching Article, What the Evidence for Learning by Doing Measures, Learning How to Learn.