Miku.

Learning by explaining, questioning, and revising your own thinking.

Miku is an experimental learning environment designed to help young learners become more aware of how they understand something, not just whether they can produce the right answer.

Instead of giving answers immediately, Miku creates situations in which learners have to explain, inspect and revise their own reasoning.

Miku, the learner's pupil

The demo

miku.app / teach / why_do_we_have_seasons
01 — Teach

You teach Miku what you learned.

02 — Build

Miku connects what you teach into a knowledge map.

03 — Predict

What does Miku know, and where might it be confused?

04 — Reflect & Re-teach

See how Miku performs, then improve its understanding.

Why Miku

AI can increasingly give students the right answer. But learning is not only about arriving at an answer.

Miku is designed to make the learner's own thinking visible:
- what they understand?
- where their explanation breaks down?
- how their mental model changes after encountering new evidence?

How learning works

Rather than immediately correcting the learner, Miku asks them to explain what they believe, challenges gaps in that explanation, and encourages them to reconstruct their understanding.

Inspired by learning by teaching

Miku thinking

Miku takes inspiration from Betty's Brain, a learning environment developed at Vanderbilt University around the idea of learning by teaching.

Students teach Betty by building a causal map of a science topic — ecosystems, climate change, thermoregulation — from a set of provided resources. They check what she has understood by asking her questions and having her take quizzes, which a mentor agent, Mr. Davis, grades and comments on. Betty only knows what the student has taught her, so a wrong quiz answer points back to the student's own map rather than to the student.

The environment is deliberately open-ended: some students model first and read later, others read first, others work resource by resource. That freedom is the point — it asks students to manage their own strategy, which is why the system is studied as a setting for self-regulated learning. It is developed by the Open-Ended Learning Environments lab at Vanderbilt University.

Miku keeps the same bargain — the agent knows only what you taught it — but replaces the causal map editor with conversation, so the explanation is produced in the learner's own words and questioned as it is given.

Theoretical framework

What the design borrows from

Four findings from the learning sciences shape how the loop is built. Each one is stated with what it implies for the design, so the choices can be argued with.

01
Explaining teaches more than reading

Students who explain material to themselves as they study learn more than students who study the same material without explaining it, largely because the attempt exposes what the explanation is missing.

In Miku, the learner does the explaining and the agent does the listening. Nothing moves forward until something has been said out loud.

Chi et al., 1989; 1994
02
Students work harder for someone else

Learners who teach a computer agent put in more effort and persist longer than learners working for their own score, and read the agent's mistakes as a problem to fix rather than as a judgment of themselves.

Miku is a pupil, not a tutor. When it gets something wrong, the question on the screen is what it was not taught.

Chase, Chin, Oppezzo & Schwartz, 2009
03
People misjudge what they know

Judgments of one's own learning are frequently wrong, and usually too confident. The mismatch is easiest to see when a prediction is made before an outcome, then held next to it.

Before the quiz, the learner predicts each question Miku will get right or wrong. Reflect shows the two side by side.

Nelson & Narens, 1990
04
Being told is not the same as changing your mind

A learner has to become dissatisfied with their current explanation before a better one can replace it. Correction supplied too early tends to be filed away rather than integrated.

Miku never corrects. The dissatisfaction comes from watching your own explanation fail a question you thought it covered.

Posner, Strike, Hewson & Gertzog, 1982
Learning hypotheses

What this prototype is testing

The core claim is narrow and falsifiable: an agent whose knowledge is strictly bounded by the learner's teaching log turns the learner's own explanation into the only variable that matters, so gaps in that explanation become visible as agent failures rather than as marks on the learner.

H1
Teaching a faithfully ignorant agent surfaces more gaps than answering questions about the same material.
MeasureConcepts a learner adds only after the agent fails a question, relative to the concepts present in the first teaching pass.
H2
Item-level prediction before the quiz improves calibration within a single session.
MeasureAgreement between per-item predictions and graded outcomes, compared against the learner's overall out-of-ten estimate.
H3
Re-teaching after a failed item produces structural revision, not simply more detail.
MeasureChanges in the extracted concept map between passes: new relations and corrected relations, not only added nodes.
H4
Learners predict taught material better than material they never taught.
MeasurePrediction accuracy on the quiz's two transfer questions, held against the eight drawn from what was directly taught.

All four depend on one condition holding: the pupil must not answer from knowledge it was never taught. If it does, a failed question no longer points back to the learner's explanation, and none of these measures mean anything.

Not another AI tutor

Miku confused

Miku is not designed to think on behalf of the learner. It is designed to create situations in which the learner has to think.

Traditional AI tutoring often follows
Student asksAI explainsStudent receives
Miku reverses that relationship
Student explainsMiku questionsStudent investigatesStudent revises
How it is built

Methodologies

The prototype is written to test the design with small scale pilot & interview. The constraints that carry the argument — the agent knows only the teaching log, predictions are made item by item, the session can be read afterwards — are enforced in code rather than left to prompting.

Interface
React · Next.js

The four screens of the loop as one piece of session state, built mobile-first because the learners are on phones and tablets.

Agents
JavaScript · model calls on the server

Two roles: the pupil, prompted only with what it has been taught, and a quizmaster that holds the answer key and never shares it with the pupil.

Checking the constraint
Untaught-question probes

A fixed set of questions the pupil was never taught, run after any change to its prompting, to check that it admits ignorance instead of answering from the model's own knowledge.

Record
One session, one file

Teaching turns, predictions and grades export together, with no accounts and nothing else stored — the working scope of a prototype used with children.

Status: a single topic, a ten-question quiz, and one loop from teaching to reflection. Everything above describes the current prototype, not a study that has been run.

What I am still exploring

  1. 01When should an AI intervene, and when should it let a learner struggle?
  2. 02How can we distinguish genuine reflection from reflective-sounding responses?
  3. 03Does repeated learning-by-teaching improve metacognitive behaviour beyond one lesson?
  4. 04How should these interactions change across different ages and subjects?
  5. 05How can we measure progress without turning metacognition into another score to optimise?
Why now

When answers are abundant and cheap, the scarce capacity is judgment about answers.

Questioning, explaining, recognising uncertainty, evaluating evidence and revising beliefs.

Generative systems have removed the friction that used to force explanation. Much of the metacognitive work of school — attempting an account, noticing it fails, repairing it — happened in that friction. If the friction is gone, the work has to be designed back in deliberately rather than assumed.

Miku is one testable instance of that design: a bounded agent, an explicit prediction, a legible knowledge structure and an exported session. Whether learning by teaching survives contact with a fluent language model is an empirical question, and this prototype exists to ask it rather than to assert an answer.

The real Miku in the passenger seat
The original Miku, April 2026.
One last thing

Meet Miku (米酷)!