Explaining It Out Loud: The Research Behind Socratic Questioning and the Feynman Technique
Published: September 17, 2026 | By: Roman Vasylyshyn, founder of Lambdio | Reading time: 9 min
There's a well-worn piece of study advice that says if you can't explain something simply, you don't really understand it. It's usually attributed to Richard Feynman, and it's the basis for the "Feynman Technique" — learn something, then explain it as if teaching a beginner, and notice where your explanation falls apart.
It's good advice. It's also, unusually for study advice, backed by a fairly deep body of experimental research — spanning three related effects that explain why producing an explanation does something that consuming one doesn't.
But there's a catch that most articles on this topic skip: the research on question-driven, "discover it yourself" learning is genuinely mixed, and some of the most influential papers in the field are sharply critical of it. The distinction between what works and what doesn't turns out to be specific and important, and it's worth getting right.
Table of Contents
- The Generation Effect
- Self-Explanation: Talking to Yourself, Productively
- The Protégé Effect: Learning by Teaching
- Where the Research Gets Critical
- What Separates Good Questioning From Bad
- Why This Pairs With Spaced Repetition
- How Lambdio Applies This
The Generation Effect
In 1978, Norman Slamecka and Peter Graf ran a deceptively simple experiment. Participants were given word pairs. One group simply read them — for example, seeing "rapid — fast." Another group had to generate the second word themselves from a cue, seeing something like "rapid — f____" and producing "fast."
The group that generated the words remembered them substantially better, despite spending no more time and seeing the same material. The finding was named the generation effect, and it's been replicated across word lists, sentences, arithmetic problems, and conceptual material in the decades since.
The mechanism overlaps with the testing effect we covered in a previous post, but it's subtly broader: it's not only about retrieving something you stored earlier, it's about the memory benefit of producing material yourself rather than receiving it. Information you construct — an answer, an example, an explanation, a connection between two ideas — is encoded more durably than the same information handed to you in finished form.
This is the most fundamental reason a conversation with a tutor tends to outperform a lecture on the same content: a conversation makes you produce things.
Self-Explanation: Talking to Yourself, Productively
A related line of work by Michelene Chi and colleagues in the late 1980s and early 1990s studied what happened when students were prompted to explain physics material to themselves as they worked through worked examples.
Students who spontaneously generated explanations — asking themselves why a step followed from the previous one, connecting it to principles they already knew — learned substantially more than students who read the same examples without explaining. When students who didn't spontaneously self-explain were simply prompted to do so, their learning improved as well, which is the important part: self-explanation isn't just a trait that good students happen to have, it's a behavior that can be induced, with measurable results.
The prompts that worked were not complicated. Asking "why does that step work?" or "how does this relate to what you just said?" at the right moments was enough. This is essentially what a Socratic dialogue does, mechanized.
The Protégé Effect: Learning by Teaching
The third strand is the most directly relevant to the Feynman Technique. Research on teachable agents — software "students" that a learner teaches — found that people put more effort into learning when they believe they're learning in order to teach someone else, and they learn more as a result. Work by Chase, Chin, Oppezzo and Schwartz (2009) named this the protégé effect: students working to teach a computer agent invested more effort and performed better than students learning the same material for themselves.
Follow-up work by Fiorella and Mayer separated two things that are easy to conflate: merely expecting to teach, versus actually teaching. Both produced benefits, but actually delivering the explanation produced the larger and more durable one. Simply telling yourself "I'll have to explain this later" helps a little. Sitting down and explaining it helps considerably more.
That's the empirical core of the Feynman Technique, and it's worth noting how the effects stack: explaining something to a listener involves generating material (generation effect), retrieving it from memory (testing effect), and articulating the connective reasoning between ideas (self-explanation) — all at once. It isn't one mechanism, it's three overlapping ones, which is likely why the technique has held up so well.
Here's roughly how those conditions compare across the literature — not from a single study, but as a representative pattern of the direction and rough magnitude these effects tend to show:
Rereading, once again, is the weakest of the active-looking options — a theme that recurs across nearly every area of learning research.
Where the Research Gets Critical
Here's the part most articles about "Socratic AI tutors" leave out.
In 2004, Richard Mayer published a paper with the pointed title "Should There Be a Three-Strikes Rule Against Pure Discovery Learning?" — reviewing decades of evidence that unguided discovery, where students are left to work out principles for themselves without direct instruction, consistently underperforms guided approaches. In 2006, Kirschner, Sweller and Clark made an even stronger version of the argument, holding that minimally guided instruction fails because it ignores how working memory and cognitive load actually operate: novices don't have the background knowledge to structure their own search through a problem space, so they flounder.
This produced a genuine, unresolved debate in the field. Hmelo-Silver, Duncan and Chinn (2007) responded that the critique applied to unguided discovery, not to well-scaffolded problem-based and inquiry learning, which does have supporting evidence. That distinction — scaffolding, not pure discovery — is roughly where the field has settled, and it's the honest version of what the research supports.
The practical upshot is that "the AI asks you questions instead of telling you things" is not automatically good pedagogy. Asking a student to derive something they have no basis to derive is one of the better-documented ways to waste their time and damage their confidence.
What Separates Good Questioning From Bad
Based on that literature, the difference between productive Socratic questioning and unproductive floundering comes down to a few specifics:
- The learner needs enough foundation to have something to reason from. Questioning works as a way to consolidate, connect, and stress-test knowledge — not as the primary delivery mechanism for material the learner has never encountered.
- Questions should target reasoning, not recall of things never taught. "Why do you think that follows?" is productive. "What do you think the term for this is?" — for a term never introduced — is a guessing game.
- The scaffolding has to fall away as competence rises. Early on, more structure and more direct explanation; later, more open questioning. Fixed difficulty serves neither the beginner nor the improving student.
- Struggle should be productive, not total. A learner who is stuck with no path forward isn't experiencing desirable difficulty; they're just stuck. Good questioning includes knowing when to give the answer.
Why This Pairs With Spaced Repetition
These effects and spaced repetition solve two different problems, which is why they're complementary rather than redundant.
Generation, self-explanation, and teaching all shape how strongly a memory gets encoded in a given session. Spaced repetition governs when you come back to it before it decays. A well-scheduled review of shallowly-encoded material is still a review of shallow material; a deep, effortful explanation session that you never revisit still falls down the forgetting curve.
In FSRS terms — the scheduling model we covered here — the quality of the session influences how much stability a successful review actually buys you, while the algorithm decides when that review happens. You want both working.
How Lambdio Applies This
This research is the basis for the conversational modes in Lambdio, and for how they're structured differently from each other:
- Socratic mode drives learning through questions rather than exposition — but it operates on chapters within a structured course, so the questioning has actual material to work against rather than asking you to invent concepts from nothing. That's a deliberate response to the Mayer/Kirschner critique above: guided questioning on scaffolded content, not open discovery.
- Feynman mode inverts the roles and asks you to explain the material back, which is the protégé effect applied directly — you produce the explanation, and the gaps surface where your explanation stalls.
- Chat Review and Standard modes sit at different points on the scaffolding spectrum, for when you want more structure or a lighter-weight pass over material you mostly know.
All of them terminate in the same place: an Again/Hard/Good/Easy rating that feeds the FSRS engine, so the depth of the session and the timing of the next one stay connected rather than being two separate systems.
If you want to try explaining a chapter back to an AI tutor and see where your explanation falls apart, courses are free to start at lambdio.com.
Sources: Slamecka & Graf (1978), "The Generation Effect," Journal of Experimental Psychology; Chi et al. (1994), "Eliciting Self-Explanations Improves Understanding," Cognitive Science; Chase, Chin, Oppezzo & Schwartz (2009), "Teachable Agents and the Protégé Effect," Journal of Science Education and Technology; Fiorella & Mayer (2013), "The Relative Benefits of Learning by Teaching and Teaching Expectancy," Contemporary Educational Psychology; Mayer (2004), American Psychologist; Kirschner, Sweller & Clark (2006), Educational Psychologist; Hmelo-Silver, Duncan & Chinn (2007), Educational Psychologist.