Last updated: 2026-09-23

F
Fundamental / general audience

Learning by Teaching, Learning by Building: Two Ways to Find Out What You Don't Actually Know

You don't really understand something until you can teach it" and "you don't really understand something until you've built it" get repeated often enough to sound like separate pieces of folk wisdom. They're the same claim, applied to two different situations. Reading a chapter, watching a lecture, or nodding along with an explanation all let a gap in your own understanding stay invisible, because nothing forces that gap to show itself. Teaching someone else and building a working version of the thing both remove that cover: an audience asks the question you didn't think to ask yourself, and a system that has to actually run refuses to accept "I sort of get it" as an answer.

The Protégé Effect: What Preparing to Teach Actually Changes

The earliest controlled study of this, Bargh and Schul's 1980 paper, found that participants preparing to teach material to someone else scored measurably higher on a later retention test than participants studying the same material for their own recall — but a second experiment in the same paper, comparing working alone, thinking aloud, and actually teaching while performing a task, found no reliable difference between those three conditions1. That mixed result matters: it's evidence against a simple "teaching is magic" story from the very first study on the topic.

Later work sharpened what the first experiment was actually picking up. Nestojko, Bui, Kornell and Bjork ran a study where one group of participants expected to teach a text passage to another student and a second group expected an ordinary test — and no one in either group ever actually taught anyone; every participant sat the same test in the end. The group that had merely expected to teach produced more complete, better-organised recall of the passage and answered more questions correctly, especially the ones covering the passage's main points2. The effect didn't require an audience to exist. It required believing one would.

Fiorella and Mayer then disentangled the two ingredients directly, comparing teaching expectancy alone, actual teaching alone, and both together. Actual teaching produced more persistent learning gains than expectancy alone, and the combination — expecting to teach, then genuinely doing it — produced the strongest and most durable effect of the three3. Put the three studies together and the honest picture is: expecting to teach changes how you study, mostly by making you organise material around what a listener would need rather than what you'd need to merely recognise it again; actually teaching adds a further, more durable benefit on top of that, most plausibly because a real listener asks a real question you didn't anticipate.

Note well. The benefit isn't just from teaching itself — merely expecting to teach, with no audience ever materialising, already measurably changes how material gets organised and recalled. Actually teaching adds a further, more durable gain on top of that.

This is the same territory Memorising vs Learning covers under "desirable difficulties" — retrieval practice, spacing, interleaving all work by forcing effortful, structure-building engagement rather than passive familiarity. Teaching, or expecting to, is a social route into the same family of effects: it forces retrieval (you have to produce the explanation, not just recognise it), forces organisation around underlying structure rather than surface order (a listener needs the why, not just the sequence you happened to read it in), and exposes exactly the kind of inert, never-applied knowledge that page describes Gick and Holyoak finding — present in memory, unavailable the moment it's actually needed.

The Same Mechanism Without an Audience: Building a Working System

A working system is a second, non-social way of forcing the same exposure. It doesn't ask you a question you didn't anticipate the way a student would; it simply fails to run, in a specific place, for a specific reason, the moment your understanding of some piece was wrong or incomplete. The physicist Richard Feynman's blackboard at Caltech, photographed after his death in 1988, carried the line "What I cannot create, I do not understand"4 — its exact origin and date are less certain than the sentiment it captures, which is precise: the test of understanding isn't being able to describe something, it's being able to construct it and watch whether the construction actually works.

Valentino Braitenberg made a research programme out of exactly this idea. His Vehicles describes a series of imagined simple machines — sensors wired directly to motors, no central controller — and shows that wiring alone produces behaviour an observer would readily describe as fear, aggression, or even preference, with no such thing built in anywhere5. The point wasn't to build a working robot for its own sake; it was to use the act of building as a way of testing a hypothesis about how much apparently complex behaviour a simple mechanism could actually produce — a question that stays comfortably vague until someone tries to build the mechanism and see. This is also the argument behind computational cognitive modelling as a research method generally: building a working model of a cognitive process and checking what it does and doesn't reproduce is a genuine way of testing a theory of that process, not merely an engineering exercise sitting alongside the theory6.

A Worked Example From This Site's Own Project Diary

This site's Finding Myself: The Journey of Building a Self-Model is a real, extended instance of the building side of this claim, not an illustration invented to fit it. That page's own recurring "Lesson" callouts are, one after another, things the project only discovered by actually running the system and watching a specific run fail — a race condition that a loosely-worded test had let pass undetected for two milestones; a safety monitor that couldn't initially tell a legitimate curated seed from a coordinated flood; a model that narrated taking an action instead of actually invoking it. None of these were found by reading the design documents more carefully; they were found because a design that looked complete on paper had to actually run, and specific gaps in the design announced themselves the moment it did. Reading the architecture page's theory first and then the journey page's history second is a compact demonstration of the whole argument on this page: the theory sounds complete right up until an attempt to build it finds the places where it wasn't.

Applying Both, Deliberately, as a Student

  • Before an exam or a viva, produce the version of your notes meant for someone who has never seen the material — not a summary for your own future skimming, an explanation for a specific imagined listener who will ask "why" and "what if" at the points where your own understanding is thinnest. Say it aloud to a real study partner if one is available; the research above shows expecting to explain does most of the work even before anyone actually listens.
  • When a concept is abstract — an algorithm, a design pattern, a piece of theory — the strongest test of whether you actually understand it is building a small, working version of it yourself, not reading a further explanation of it or watching someone else build it. A ten-line implementation that runs on a deliberately tricky input tells you more about a genuine gap in your understanding than another paragraph of prose ever will.
  • Treat a failure in either mode as the useful event, not an embarrassing one. A question you can't answer while teaching, or a bug your small implementation produces, is exactly the gap that silent reading would have let stay invisible — that's the mechanism working, not a sign you weren't ready to teach or build yet.

Both routes are ways of doing what Pedagogy, Andragogy, Heutagogy, Rhizome calls testing capability rather than competency — handling a demand on your knowledge that wasn't scripted in advance, which is precisely what a listener's unplanned question and a program's actual runtime behaviour both deliver. And treating "did explaining or building this expose a gap?" as a routine question to ask yourself is a concrete version of the Level 5 habit Personal Learning Maturity describes: experimenting on your own study process rather than assuming it's already working.

References

  1. Bargh, J. A., & Schul, Y. (1980). On the cognitive benefits of teaching. Journal of Educational Psychology, 72(5), 593–604. https://doi.org/10.1037/0022-0663.72.5.593 ↩

  2. Nestojko, J. F., Bui, D. C., Kornell, N., & Bjork, E. L. (2014). Expecting to teach enhances learning and organization of knowledge in free recall of text passages. Memory & Cognition, 42(7), 1038–1048. https://doi.org/10.3758/s13421-014-0416-z ↩

  3. Fiorella, L., & Mayer, R. E. (2013). The relative benefits of learning by teaching and teaching expectancy. Contemporary Educational Psychology, 38(4), 281–288. https://doi.org/10.1016/j.cedpsych.2013.06.001 ↩

  4. Photographed on Richard Feynman's office blackboard at Caltech after his death in February 1988; widely reproduced, including in Caltech's own archives, though its exact date of writing is not independently documented. Treated here as a well-attested attribution, not a dated published quotation. ↩

  5. Braitenberg, V. (1984). Vehicles: Experiments in Synthetic Psychology. MIT Press. ↩

  6. Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences, 40, e253. As cited on this site's A Parallel-Drafts Architecture for Modelling the Self. ↩