The Learning by Teaching App That Puts You in the Driver's Seat
Most AI tutors give you answers. SciCirc takes the opposite approach: it asks you questions, listens to your explanations, and uses follow-up probing to reveal exactly where your understanding breaks down. This is the Feynman Technique, operationalised in software — and the research behind it is more compelling than anything in the current edtech landscape.
What is SciCirc?
SciCirc is an AI-powered learning platform built by Tsurukame Portal, a Japan-based edtech company. Its defining characteristic is deceptively simple: instead of the AI teaching the student, the student teaches the AI.
In a typical SciCirc session, the student selects a concept — say, Newton's third law, or the meaning of "opportunity cost," or how to factorise a quadratic. The AI then plays the role of a curious but slightly confused learner. It asks the student to explain the concept in plain language. As the student explains, the AI poses follow-up questions: "What does that word mean exactly?" "Why does that happen?" "Can you give me an example?" The student must answer — or realise, in real time, that they cannot.
That moment of realisation — when you discover a gap in your own understanding while trying to explain something — is precisely the moment when lasting learning happens. SciCirc is engineered to create that moment, reliably and at scale.
The Science: Bloom's 2-Sigma and the Protege Effect
The Learning by Teaching methodology is not a motivational metaphor. It is grounded in several decades of cognitive science and educational research.
Bloom's 2-Sigma Problem (1984)
In a landmark 1984 paper published in Educational Researcher, Benjamin Bloom reported results that the educational community has spent four decades trying to replicate and explain. Bloom compared three modes of instruction: conventional classroom teaching (one teacher, many students), mastery learning (structured self-paced study with feedback), and individual one-on-one tutoring. The result was striking.
Students who received individual one-on-one tutoring performed two standard deviations above students in conventional classrooms. In practical terms, the average tutored student outperformed 98% of conventionally taught students. Bloom called this "the most striking finding" of his career.
The problem, which Bloom himself posed as a research challenge, was how to replicate this effect without the impossible cost of assigning a human tutor to every student. SciCirc's answer is: make the AI the student, not the tutor. When a learner is responsible for explaining a concept clearly enough for another party to understand it, they naturally engage in the same metacognitive processes that make one-on-one tutoring so effective — without needing a trained human on the other end.
The Protege Effect
A complementary line of research, often called the Protege Effect, examines what happens when students are told they will need to teach material to someone else. In a 2011 study by Nestojko et al. (later published in Memory & Cognition), students who were told they would teach a text to another student recalled significantly more of that text — and organised their recall more systematically — than students who expected a standard test. The mere expectation of teaching changed how they processed the material.
SciCirc does not simulate the expectation of teaching. It enacts it. Every session is, by design, a teaching session. The Protege Effect is not a happy side-effect; it is the core mechanism.
Learners who prepare to teach material recall it with greater accuracy and structural coherence than those who prepare to be tested — even when no actual teaching takes place. The anticipation of explanation alone reshapes encoding. (Nestojko et al., 2014, Memory & Cognition)
The Feynman Technique as Software
Richard Feynman, the Nobel Prize-winning physicist, described his personal learning method in terms that any teacher would recognise: to truly understand something, you must be able to explain it simply enough that a child could follow along. If you cannot, you have not understood it — you have only memorised it.
"If you can't explain it simply, you don't understand it well enough."
The Feynman Technique, as it became popularised, consists of four steps: choose a concept, explain it as if to a complete beginner, identify the gaps where your explanation breaks down, and revisit the source material to fill those gaps. It is powerful precisely because it forces metacognitive honesty — you cannot fake understanding when you are required to produce a clear explanation.
The problem with applying the Feynman Technique as a solo practice is that it is easy to unconsciously lower the bar. When you are talking to yourself, you tend to gloss over the parts you do not understand, because there is no-one to ask "But wait, why does that happen?" SciCirc provides that interlocutor. The AI's follow-up questions are calibrated to probe exactly the explanatory depth that the Feynman Technique demands — and they are tailored to the concept being taught, not generic.
In this sense, SciCirc is the Feynman Technique with a patient, infinitely curious conversation partner who never lets you off the hook.
How SciCirc Differs from ChatGPT, Khanmigo, and Quizlet
The AI tutoring space has expanded rapidly since 2023. Tools like Khanmigo (Khan Academy's GPT-4-powered tutor), Quizlet's AI features, and direct use of ChatGPT are now common in classrooms and study sessions. Each represents a genuine advance over passive video lectures. But each also shares a common limitation: they keep the student in the receiving role.
| Feature | ChatGPT | Khanmigo | Quizlet AI | SciCirc |
|---|---|---|---|---|
| Student teaches the AI | ✕ | ✕ | ✕ | ✓ Core loop |
| AI asks probing follow-ups | Rarely | Sometimes | ✕ | ✓ Always |
| Explanation-quality feedback | ✕ | ✕ | ✕ | ✓ |
| Feynman Technique built in | ✕ | ✕ | ✕ | ✓ |
| Curriculum-aligned content | ✕ | ✓ | Partial | ✓ |
| Primary student role | Questioner | Questioner | Reviewer | Teacher |
Khanmigo is arguably the most thoughtful of the conventional AI tutors. It is designed not to simply give answers but to guide students toward answers with Socratic prompting. That is a meaningful step forward. But even in Khanmigo, the student remains the one asking questions — which means the cognitive load of explanation never falls on the learner.
Quizlet's AI features are primarily optimised for recall: generating flashcards, practice tests, and study guides from uploaded material. These are useful for exam preparation but do not develop the kind of conceptual flexibility that comes from being forced to articulate understanding in your own words, under pressure from a curious interlocutor.
The distinction is not trivial. Recall and understanding are measurably different cognitive capacities. Bloom's taxonomy, the same framework that produced the 2-sigma research, places "remembering" at the base of its hierarchy and "explaining" several levels higher. SciCirc operates at the explanation level from the first minute of every session.
How SciCirc Actually Works
A SciCirc session proceeds through three phases, each designed to maximise the depth of explanatory engagement.
Phase 1: Concept Selection
The student selects a concept from SciCirc's curriculum-aligned topic library — or enters a custom topic. The platform covers Mathematics (arithmetic through calculus-level reasoning), English (grammar, reading, vocabulary, writing), Science (physics, chemistry, biology, earth science), and Social Studies (history, geography, civics, economics).
Concept selection can be driven by upcoming exam topics, teacher assignments, or personal curiosity. There is no prerequisite and no onboarding friction.
Phase 2: Teaching Session
Once a concept is selected, the AI adopts the persona of a motivated but unfamiliar learner. It opens with something like: "I've heard of [concept] but I'm not really sure what it means. Can you explain it to me?"
The student explains. The AI responds with genuine engagement: it asks clarifying questions if the explanation is vague, requests examples if the concept is abstract, and occasionally offers a plausible-but-incorrect interpretation and asks whether it understood correctly. This last technique — deliberately misunderstanding to test whether the student can correct the error — is particularly effective at revealing shallow knowledge.
The session continues until the student has produced a complete, coherent explanation. This typically takes five to fifteen minutes, depending on concept complexity.
Phase 3: Feedback and Review
After the session, SciCirc analyses the student's explanation and generates structured feedback: which components of the concept were explained well, which were glossed over, which were explained incorrectly, and what the student should review next. This feedback is specific — not generic encouragement — and is linked directly to moments in the conversation where the explanation was weak.
The result is a learning loop that is both more engaging than passive review and more diagnostically precise than a multiple-choice test.
Subjects and Curriculum Alignment
SciCirc's content is aligned to Japanese national curriculum standards, which are among the most rigorously structured in the world. This means the topic library reflects coherent progressions — not just loose collections of facts — and the AI's follow-up questioning is calibrated to what students at each level are expected to understand.
- Mathematics: Number sense, algebra, geometry, probability and statistics, and mathematical reasoning. Content spans elementary through senior high school level.
- English: Grammar (parts of speech, sentence structure, tense), reading comprehension, vocabulary in context, and writing conventions.
- Science: Physics (mechanics, electricity, waves), chemistry (atomic structure, reactions, periodic table), biology (cells, genetics, ecology), and earth science.
- Social Studies: Japanese and world history, physical and human geography, political science, and basic economics.
An English-language curriculum layer is in active development. International learners can currently access the platform and engage with concepts directly, even if some interface elements remain in Japanese.
What a Session Looks Like
The following illustrates a representative SciCirc session on the concept of "photosynthesis."
[Demo screenshot — full interface available at tsurukamezan.net/scicirc/]
In a typical session on photosynthesis, the exchange might proceed as follows:
AI: "I keep hearing the word 'photosynthesis' in biology class but I don't really get it. Can you explain what it actually is?"
Student: "Photosynthesis is how plants make food using sunlight."
AI: "Okay, so plants just... absorb sunlight and that becomes food? What does 'food' mean for a plant exactly? Is it the same as what we eat?"
This exchange already does something a flashcard never could: it forces the student to confront the ambiguity in their initial explanation. What is "food" for a plant? The student must now explain glucose, or admit they do not know. Either way, they have identified a gap — which is precisely the goal.
By the end of the session, the student has been walked through chlorophyll, carbon dioxide, water, the light-dependent and light-independent reactions — not by reading a textbook, but by having to produce an explanation clear enough to satisfy a curious interlocutor.
A Note for Educators and Researchers
SciCirc is not a toy and it is not a productivity application. It is a deliberate implementation of a specific pedagogical theory, built by someone who has spent significant time studying the gap between what educational technology promises and what cognitive science actually supports.
If you are an educator interested in piloting SciCirc with a class, or a researcher interested in studying its effects on learning outcomes, we would welcome the conversation. Contact us here.
We are particularly interested in collaboration with researchers working on:
- AI-mediated Socratic dialogue and its effect on conceptual understanding
- The Protege Effect in digital learning environments
- Comparative studies of explanation-based versus recall-based AI tutoring
- Curriculum alignment and adaptive learning in East Asian educational contexts