The protégé effect is the finding that you learn material better when you study it in order to teach it. It is real and well replicated. The popular version hides the part that decides how much you actually get: how you teach matters more than that you teach. Teaching a person who can interrupt you produced an effect of g = 0.84 in Keiichi Kobayashi's meta-analysis, while explaining into a recording with no advance warning produced g = 0.23.
What is the protégé effect?
It is the gap between studying something for yourself and studying it to pass on. The name comes from research on teachable agents, software characters that students instruct and that then reason out loud from whatever they were taught.
The founding study is Chase, Chin, Oppezzo and Schwartz, published in the Journal of Science Education and Technology in 2009. Eighth graders worked through the same biology software. One group believed they were teaching a character called Betty. The other believed they were learning for themselves. The teaching group spent more time reading and scored higher, and the benefit was largest for the lower achieving students.
So the effect was first measured with a piece of software as the pupil, which matters later.
How strong is the effect, and does teaching have to be real?
Kobayashi has published two syntheses, and read together they answer both halves.
The first, in Japanese Psychological Research in 2019, pooled 28 studies. Preparing to teach and never teaching produced g = 0.35. Preparing to teach and then doing it produced g = 0.56. Both held up on delayed tests, and both improved deep understanding rather than only recall.
The second, in Frontiers in Psychology the same year, split those conditions by how much the teaching talks back:
| What you did | Effect (g) |
|---|---|
| Explained into a video or text, no preparation | 0.23 |
| Prepared for a video or written explanation | 0.27 |
| Prepared to teach a person face to face | 0.50 |
| Gave a video or written explanation after preparing | 0.48 |
| Taught a person face to face after preparing | 0.84 |
Read the first and last rows together. Both are teaching. One is worth roughly three and a half times the other. The variable is whether the thing you are explaining to can respond.
That table is the practical heart of this research, and it is almost never the part that gets repeated.
Why does a listener who answers beat a notebook?
Because a notebook accepts anything. You can write a confident sentence about a concept you have not understood, and nothing pushes back. Every gap survives in writing, which is the same reason highlighting feels productive and changes nothing.
A listener who can answer removes that option. They ask which part you meant. They repeat your explanation back slightly wrong. They ask the obvious follow-up you had quietly skipped. Each of those forces a second pass over the exact place your understanding was thin, and you cannot choose where that place is.
Kobayashi's own reading of the gap is that non-interactive explaining lets people process the material shallowly and still finish the task. Finishing is the problem. An interactive partner makes finishing conditional on being understood.
Does it work if you only expect to teach?
Yes, and this is the cheapest version of the technique. Nestojko, Bui, Kornell and Bjork tested it in Memory & Cognition in 2014. One group read a passage expecting a test, another read it expecting to teach it. Nobody taught anyone. Everybody was tested.
The group that expected to teach recalled more, organized their recall better, and did best on the main points rather than on trivia. The expectation alone changed how they read.
But the timing is not negotiable. Kobayashi found that the benefit depends on people knowing in advance. Deciding to explain something after you have already studied it does not produce the effect, because the effect happens during the studying. This is why "I will summarize it afterwards" is a weaker plan than it sounds.
How do you use this on a podcast or an audiobook?
Spoken material makes both halves of the finding harder. You cannot skim back over a paragraph, and you usually have your hands full, so the notebook option is mostly gone anyway.
The expectation half still works, and it is free. Decide before you press play that you are going to explain this episode to someone specific. Not "I will pay attention", but a named person and a real moment: your partner at dinner, a colleague on Thursday. That is the Nestojko condition, set at the only time it can be set.
The interactive half is the harder one to arrange, because it needs a listener who is available when you finish, and who will keep asking. A few ways people solve it:
- A standing conversation with one person. The strongest version, and the hardest to schedule.
- A voice memo addressed to someone. Better than notes, still no one answering. Roughly the 0.48 row.
- Talking to an AI about the episode. A listener that asks back, available at the end of the drive.
That last one is what Graspd is built around, so treat the next paragraph as an interested party talking. The companion only ever sees the transcript up to where you have listened, so it can ask about what you just heard without telling you what happens later in the book. You explain the idea, it asks the follow-up, and the gaps show up in the answer you could not give. Those are the gaps that an illusion of competence hides while the episode is still playing. Marks you drop while listening give the conversation something to start from, which is the same loop described in how to remember what you listen to.
Two honest limits. Kobayashi's 0.84 row was measured with human partners, and Betty's Brain was measured on eighth graders doing science, so nobody has tested whether an AI listener reaches the face to face number for an adult on a podcast. And Graspd cannot import Audible files, because they carry DRM, so audiobooks have to come from a source that gives you the actual file.
What do the popular explanations get wrong?
Two things, and both are checkable.
The most visible explainer for this term is an Annie Murphy Paul piece in Psychology Today from June 13, 2012. It is a good article. It is also fourteen years old, which puts it before both Kobayashi syntheses and before the Nestojko study, so it cannot contain the numbers that tell you which version of the technique to use. It leans instead on the finding that firstborn children score higher on intelligence tests, attributed to time spent instructing younger siblings, which is a correlation carrying a lot of weight.
The second is the repeated advice to "explain it to someone" with no mention that the effect roughly triples when that someone can interrupt. If you take the common version of this tip, you will reach for a journal or a voice memo, because those are easiest. Those are the bottom rows of the table.
There is a third pattern worth naming. A bought summary is the inverse of this entire finding: someone else did the explaining, and the work that produces the effect has already been done for you by a stranger. That is the real cost hiding in the summary subscriptions, and it applies whether the summary is good or bad.
If you want the version of this that needs no listener at all, testing yourself gets at the same gaps from a different direction, and quizzing yourself on podcasts is the closest neighbor to teaching. For text rather than audio, the same ordering of techniques shows up in how to remember what you read.
FAQ
Is the protégé effect the same as the Feynman technique?
They overlap but are not identical. The Feynman technique is a procedure: explain a concept in simple language, find where you stall, go back. The protégé effect is the measured finding that expecting to teach changes how you study. Feynman is one way to produce the effect, and it is the non-interactive way.
Does the protégé effect work with an AI instead of a person?
Partly proven. The founding 2009 study used a software pupil and found the effect, so a non-human listener can produce it. But the largest measured effect came from face to face teaching with human partners, and no study has tested a modern AI listener against that number. Treat it as promising rather than settled.
Do I have to actually teach, or is planning to teach enough?
Planning alone works. Pooled across 28 studies, preparing to teach without teaching produced an effect of g = 0.35, versus g = 0.56 when people went on to teach. So the intention gets you most of the way, and following through adds the rest. The intention has to come first.
Why does it not work if I decide to explain it afterwards?
Because the effect happens while you study, not while you talk. Knowing you will teach makes you hunt for main points and build a structure as you go. If that instruction arrives after you have finished reading or listening, the studying already happened in the ordinary way, and there is nothing left to change.
What is the easiest way to start?
Before you press play, pick one real person and one real occasion for explaining the episode to them. That single sentence is the whole Nestojko condition and costs nothing. Then find any listener who will ask you a follow-up question, because that is where most of the measured benefit sits.
