Post
JA EN

Teaching AI as a Way to Learn — The Protégé Effect and the Inverted Logic of Education

Posted
By
18 min read
Teaching AI as a Way to Learn — The Protégé Effect and the Inverted Logic of Education
  • Target audience: Educators, knowledge workers interested in AI, managers
  • Prerequisites: Basic experience using AI tools (ChatGPT, Claude, or similar)
  • Reading time: 15 minutes

Overview

“If AI can teach us anything, will human-to-human teaching become obsolete?” — This concern comes up often in conversations about education in the AI era.

For now, the worry is probably unfounded. A Microsoft Research study (Baym et al., 2026, n=557) found that 88% of high-performing AI users cited “informal conversations with colleagues” as central to their AI proficiency1. Sharing prompt tips over coffee, swapping failure stories at lunch — people are actually teaching each other how to use AI more, not less.

But what if AI becomes capable enough to handle that meta-level teaching too? Will “learning by teaching” disappear then?

That question has a crucial blind spot. Teaching is not only learning for the student — it is learning for the teacher. And AI may actually be a better recipient of teaching than any human.

In psychology, the phenomenon where teachers learn more than their students is known as the protégé effect. Chase et al. (2009), who coined the term, demonstrated that students who taught a virtual agent invested significantly more effort and learned more deeply than students studying for themselves2. A meta-analysis by Kobayashi (2019) confirmed the effect at Hedges’ g = 0.56 — a medium-to-large effect size3.

From 2024 onward, a wave of research has begun reimagining this “learning by teaching” in the context of AI. Studies on Teachable Agents powered by LLMs — systems where the AI plays the learner — are showing meaningful gains in programming education4 and mathematics5. More striking still is the observation that everyday AI use — designing prompts, writing AI skill instructions — may be unconsciously reproducing the same cognitive processes as formal teaching.

This article traces the protégé effect from its academic foundations, examines why instructing AI shares the same cognitive mechanism as the act of teaching, and makes the case for an inverted educational logic: not “letting AI teach us,” but “learning by teaching AI.”

The Protégé Effect — The Science of “Teaching as Learning”

The Discovery

The intuition that “teaching something is the best way to learn it” is ancient. Scientific evidence for it, however, is relatively recent.

Chase et al. (2009) demonstrated the effect using the Betty’s Brain system at Vanderbilt University2. Betty’s Brain is a software environment where students teach a virtual character named “Betty” by building concept maps. Betty uses the knowledge she receives to perform qualitative reasoning and answer questions.

The experiment divided eighth-graders into two groups. Both used the same software, but one group believed they were “teaching Betty” while the other believed they were “studying for themselves.” The results were unambiguous:

  • The Teachable Agent (TA) group spent significantly more time on learning activities
  • The TA group learned more — an effect especially pronounced for lower-achieving students
  • In a follow-up experiment with fifth-graders, students displayed social behavior toward Betty: attributing mental states to her, feeling responsible for her failures, and expressing negative emotions when she underperformed

In short, the sense of responsibility that comes from “teaching for someone else” amplified both learning motivation and learning behavior.

Effect Sizes from Meta-Analysis

Kobayashi (2019) quantified this effect in a meta-analysis synthesizing 28 studies3. The results revealed that interactivity is the key determinant:

ConditionEffect size (Hedges’ g)
Preparation to teach only0.35
Preparation + actual teaching0.56
Direct face-to-face teaching + preparation0.84
Expectation of indirect teaching (e.g., video creation)0.27
Expectation of direct teaching0.50

Notably, even preparing to teach has an effect — but actually teaching produces a larger one. And the more interactive the exchange — the more real-time feedback comes back from the learner — the greater the learning gain.

Why Teaching Produces Learning — Four Mechanisms

Research by Fiorella & Mayer (2013, 2014)67 and peer tutoring studies by Roscoe & Chi (2007, 2008)89 converge on four mechanisms:

flowchart TB
    T["The Act of Teaching"] --> M1["1. Metacognitive Activation"]
    T --> M2["2. Knowledge Structuring"]
    T --> M3["3. Generation Effect"]
    T --> M4["4. Motivational Enhancement"]
    M1 --> R["Deep Learning"]
    M2 --> R
    M3 --> R
    M4 --> R

1. Metacognitive Activation

People preparing to teach use 1.3× more metacognitive strategies than those studying without the expectation of teaching (Muis et al., 2015)10. Teaching requires constant self-monitoring: “Do I actually understand this well enough to explain it?” This self-checking creates opportunities to detect and repair gaps in understanding.

2. Knowledge Structuring

Roscoe & Chi (2007) found that the act of teaching operates in two modes: “knowledge-telling” (summarizing material) and “knowledge-building” (reconstructing one’s own understanding and using inference to fill gaps)8. Deep learning occurs in the second mode — not when you simply recap what you read, but when you actively reorganize your mental model.

3. Generation Effect

Teaching requires actively retrieving information from memory and reconstructing it in your own words. This generative process leaves stronger memory traces than passive reading. Fiorella & Mayer (2014) confirmed that students who actually taught others performed best on delayed retention tests7 — the generation effect is particularly valuable for long-term memory.

4. Motivational Enhancement

As Chase et al.’s experiment showed, the social responsibility of “teaching for someone” elevates intrinsic motivation2. You might cut corners when studying for yourself; you cannot cut corners when someone is counting on you. This psychology drives deeper engagement with the material.

AI as a Learner — Three Advantages Over Human Students

From Betty’s Brain to LLMs — The Evolution of Teachable Agents

Protégé effect research reached a turning point with the emergence of LLMs in the 2020s.

Earlier Teachable Agents like Betty’s Brain relied on structured interfaces — students entered knowledge through concept maps. The TeachYou system, presented at the CHI 2024 conference, introduced a new approach: using an LLM as the learner4. A chatbot named AlgoBot plays the role of someone who “doesn’t know,” actively asking “Why?” and “How does that work?” to probe the student’s explanations. Xing et al. (2025)’s ALTER-Math system applied the same approach to mathematics education, finding significant knowledge gains over a control group in a study of 320 middle school students5.

But the point I want to emphasize is this: intentionally designed Teachable Agents are not the only “learners” available to us. The AI assistants we use every day share the same fundamental structure.

Why AI Is a More Rewarding Student Than Any Human

As Kobayashi (2019)’s meta-analysis showed, learning gains scale with interactivity3. AI interaction mirrors the structure of face-to-face teaching while adding unique advantages of its own:

Radical Honesty — No Polite Pretending

A human student will often infer what the teacher wants and pretend to understand. AI exposes ambiguity without mercy. You think you explained something clearly — if the instructions were vague, AI simply won’t perform as expected. This mismatch acts as a mirror, reflecting gaps in your own understanding.

In programming, “rubber duck debugging” is well-known — explaining your code to a rubber duck helps you spot bugs11. AI is an evolution of the rubber duck. It does not just “listen” to your explanation; it asks follow-up questions and acts on its (mis)understanding, surfacing implicit assumptions far more effectively.

High-Speed Feedback Loops

In human-to-human teaching, you often don’t know whether your explanation was correct until a test or a real-world outcome reveals it. AI delivers the result of your instructions immediately. This speed of feedback promotes what Roscoe & Chi (2008) called “reflective knowledge-building”9. The improvement cycle that a human teacher accumulates over a semester plays out dozens of times in a single day with AI.

Psychological Safety

Making a mistake in front of a human student is embarrassing. That psychological cost creates a barrier to entry for “learning by teaching” — especially for beginners. Chase et al. (2009) found that the protégé effect was strongest for lower-achieving students2. When your “student” is an AI, the social cost of failure is zero. This safety fulfills the autonomy and competence needs described in Deci & Ryan (1985)’s self-determination theory12, removing barriers to engagement.

This is also a natural implementation of the Feynman Technique. Physicist Richard Feynman’s learning method — explain a concept as if to a twelve-year-old; whenever you get stuck, go back and relearn — is reproduced organically through dialogue with AI.

The Learning Loop — Teaching and Being Taught in Alternation

The Limits of the One-Way Model

So far, we have examined the theory behind the protégé effect and why AI makes an excellent learner. But real AI use is not a one-way street in which humans teach AI. More accurately, it is a cycle where teaching and learning alternate:

Imagine asking AI to research a topic:

  1. AI researches and presents information → You gain new knowledge (learning phase)
  2. You review the output and give feedback → “This perspective is missing,” “This assumption is off” (teaching phase)
  3. AI returns an improved output → You gain higher-quality knowledge (learning phase)
  4. You return still more precise feedback → Your own understanding deepens (teaching phase)
flowchart TB
    A["AI Researches and Delivers"] --> B["Human Gains Knowledge<br>(Learning Phase)"]
    B --> C["Human Reviews and<br>Gives Feedback<br>(Teaching Phase)"]
    C --> D["AI Improves"]
    D --> A
    
    B -.->|"Knowledge Acquisition"| L["Learning"]
    C -.->|"Protégé Effect:<br>Metacognition, Structuring,<br>Generation Effect"| L

The key insight is that both phases produce learning. The learning phase brings new knowledge; the teaching phase deepens understanding through the protégé effect. And with each revolution of the cycle, both the quality of what you receive and the precision of your feedback improve.

This structure is fundamentally different from the one-way “AI teaches → human learns” model. When an AI tutor only transmits knowledge, the learner is a receiver. But when you “teach back” through review and feedback, you become an active constructor of knowledge. As Kobayashi (2019)’s meta-analysis demonstrates, this active role is what drives the difference in learning outcomes — preparation only (g = 0.35) versus actually teaching (g = 0.56)3.

The Four Mechanisms Appearing in the Feedback Phase

When you dissect the “teaching phase” of this cycle, the four mechanisms of the protégé effect appear directly.

Prompt Design — Immediate Teaching-as-Learning

What happens when you write a prompt?

You type “List review perspectives.” The result is not what you expected. You realize “review perspectives” was ambiguous. So what exactly do you need? Code quality? Security? Performance? You re-articulate what you actually want.

MechanismPrompt design equivalent
Metacognitive activationSelf-monitoring: “What exactly do I want here?”
Knowledge structuringConverting a vague need into precise, structured instructions
Generation effectTranslating tacit understanding into explicit, language-based instructions
Motivational enhancementImmediate feedback through actual task execution

A study published in Frontiers in Education showed that the essence of prompt engineering is the ability to clearly articulate a problem, its context, and its constraints13. That definition maps directly onto the definition of teaching ability. The same study notes that writing prompts inherently demands metacognitive self-monitoring13.

AI Skill Design — Teaching That Converts Tacit Knowledge to Explicit

A deeper form of “teaching” than individual prompts is designing AI “skills” or “rules” — comprehensive instruction sets for performing a specific task. Writing one is analogous to designing a curriculum:

  • Systematizing domain knowledge: What to convey, in what sequence, at what level of granularity
  • Anticipating edge cases: Thinking ahead to “this instruction will break down in situation X”
  • Articulating tacit knowledge: Writing out the things you “just do this way” without having formalized why

Nonaka & Takeuchi (1995) called the conversion of tacit knowledge into explicit knowledge “externalization”14. As Polanyi (1966) wrote, “We can know more than we can tell”15 — this conversion is inherently difficult. But designing AI skills forces it. AI does not understand implicit context. “Just do it right” will not work. You must translate your expertise into a form that another intelligence can understand.

The Core of AI Literacy Is the Ability to Teach

What we call “AI literacy,” then, is not primarily the ability to receive AI outputs — it is the ability to return precise feedback to AI, to teach. And “the ability to teach” grows autonomously within this cycle.

Practical Implications — Working with the Cycle in Mind

For Individuals: Do Not Stop at the Learning Phase

To convert AI interactions into genuine learning, consciously build the “teaching phase” into the cycle:

1. Grade AI outputs like exam papers

When you receive an output, do not just use it. Examine it the way a teacher grades an exam. When you find errors or missing perspectives, explain to the AI why it falls short. This “grader” stance activates critical thinking and metacognition.

2. Teach back what you just learned

After learning a new concept from AI, try explaining it back in your own words. Start a conversation with a setup like: “Assume you know nothing about this topic. I’ll explain it — please ask questions if something is unclear.” Teaching immediately after learning starts the cycle spinning at once.

3. Write your own AI skills and prompt templates

For tasks you repeat regularly, try writing instructions that delegate the task to AI. The act of writing often clarifies, for the first time, “what standard do I actually use to make this judgment?” Then review the results and refine the instructions — that is the cycle itself.

For Educators: Make AI the Student

Traditional assumption: AI teaches in place of the teacher → teachers become unnecessary

Inverted logic: AI plays the student → students learn by teaching

As research on TeachYou4 and ALTER-Math5 shows, using an LLM as “the one being taught” is technically feasible and educationally validated.

  • Comprehension check: “Can the student teach this to AI?” as a test of understanding
  • Explanation practice: When AI says “I don’t understand,” the student practices rephrasing in simpler terms
  • Experiencing the cycle: Designing assignments around the full sequence of having AI research a topic, reviewing the output, and feeding back

For Organizations: Making Tacit Knowledge Tangible

Designing AI skills is also a powerful mechanism for converting organizational tacit knowledge into explicit form. Having experienced employees articulate their “I just do it this way” intuitions as AI instructions is exactly what Nonaka & Takeuchi (1995) called “externalization”14.

Traditional knowledge management — writing manuals, building knowledge bases — suffers from weak motivation. Write a document and no one may ever read it. But when you articulate expertise as an AI skill, AI uses it immediately. This instant feedback fundamentally changes the incentive to articulate knowledge.

Limitations and Caveats

Limitations of the Protégé Effect Research

The studies cited in this article have the following limitations:

  • Chase et al. (2009)’s experiment used middle school students, and generalizing to adult learning requires caution2
  • The effect size (g = 0.56) in Kobayashi (2019)’s meta-analysis is an aggregate of 28 studies, and individual study results vary considerably3
  • LLM-based Teachable Agent research (TeachYou, ALTER-Math) consists entirely of studies published in 2024–2025, and long-term learning effects remain unverified45

The Analogy Between AI Instruction Design and Teaching Has Limits

The central claim of this article — that designing AI instructions reproduces the protégé effect — is theoretical reasoning based on mechanistic similarity, not a direct experimental finding. Whether prompt design or skill creation produces learning gains equivalent to classroom teaching remains an open empirical question.

The Difference Between “Teaching” and “Instructing”

Not all AI use qualifies as “learning by teaching.” Simple commands like “translate this” or “summarize this” correspond to what Roscoe & Chi (2007) called “knowledge-telling”8, and are unlikely to produce deep learning. To gain a learning benefit, the instruction design must be at a level that involves reconstructing (knowledge-building) your own understanding.

Conclusion

“In an era when AI can teach us anything, teaching each other will become obsolete” — this concern overlooks two things. First, the new teaching of how to use AI is actually increasing1. More fundamentally, as protégé effect research shows, the recipient of teaching need not be human. Teaching AI is itself a form of learning.

The depth of learning is determined not by how you receive information, but by how actively you construct knowledge. Teaching activates four mechanisms — metacognitive activation, knowledge structuring, the generation effect, and motivational enhancement — that promote this active construction.

And AI is an ideal recipient for this kind of learning. It exposes ambiguity without social grace, returns feedback at high speed, and never judges you for failing. Everyday prompt design and AI skill creation reproduce the cognitive benefits of the teaching act, whether we intend them to or not.

What AI-era education truly needs may not be “how to make AI teach better” — but learning how to teach AI well.

Other articles on related themes:

References

References are listed in the order they appear in the text.

Additional References (not directly cited in text)

  1. Peer Influence Can Make or Break Your AI Rollout - Baym, N., Dillon, E., & Jaffe, S. Harvard Business Review (2026). Microsoft Research survey, n=557. [Reliability: Medium–High] ↩︎ ↩︎2

  2. Teachable Agents and the Protégé Effect: Increasing the Effort Towards Learning - Chase, C. C., Chin, D. B., Oppezzo, M. A., & Schwartz, D. L. Journal of Science Education and Technology, 18(4), 334-352 (2009). [Reliability: High] ↩︎ ↩︎2 ↩︎3 ↩︎4 ↩︎5

  3. Interactivity: A Potential Determinant of Learning by Preparing to Teach and Teaching - Kobayashi, K. Frontiers in Psychology, 9, 2755 (2019). Meta-analysis (28 studies). [Reliability: High] ↩︎ ↩︎2 ↩︎3 ↩︎4 ↩︎5

  4. Teach AI How to Code: Using Large Language Models as Teachable Agents for Programming Education - CHI Conference on Human Factors in Computing Systems (2024). Peer-reviewed conference paper. [Reliability: High] ↩︎ ↩︎2 ↩︎3 ↩︎4

  5. Development of a generative AI-powered teachable agent for middle school mathematics learning - Xing, W. et al. British Journal of Educational Technology, 56(5), 2043-2077 (2025). n=320, peer-reviewed. [Reliability: High] ↩︎ ↩︎2 ↩︎3 ↩︎4

  6. The relative benefits of learning by teaching and teaching expectancy - Fiorella, L. & Mayer, R. E. Contemporary Educational Psychology, 38, 281-288 (2013). Peer-reviewed. [Reliability: High] ↩︎

  7. Role of expectations and explanations in learning by teaching - Fiorella, L. & Mayer, R. E. Contemporary Educational Psychology, 39, 75-85 (2014). Peer-reviewed. [Reliability: High] ↩︎ ↩︎2

  8. Understanding Tutor Learning: Knowledge-Building and Knowledge-Telling in Peer Tutors’ Explanations and Questions - Roscoe, R. D. & Chi, M. T. H. Review of Educational Research, 77, 534-574 (2007). Peer-reviewed review article. [Reliability: High] ↩︎ ↩︎2 ↩︎3

  9. Tutor learning: the role of explaining and responding to questions - Roscoe, R. D. & Chi, M. T. H. Instructional Science, 36, 321-350 (2008). Peer-reviewed. [Reliability: High] ↩︎ ↩︎2

  10. Learning by preparing to teach: Fostering self-regulatory processes and achievement during complex mathematics problem solving - Muis, K. R., Psaradellis, C., Chevrier, M., Di Leo, I., & Lajoie, S. P. Journal of Educational Psychology, 107(4), 1-19 (2015). Peer-reviewed. [Reliability: High] ↩︎

  11. Rubber Duck Debugging: The Psychology of How it Works - ThoughtfulCode. Explanation of cognitive activation through verbalization. [Reliability: Medium] ↩︎

  12. Self-Determination and Intrinsic Motivation in Human Behavior - Deci, E. L. & Ryan, R. M. Springer (1985). Foundational work in self-determination theory. [Reliability: High] ↩︎

  13. Prompt engineering as a new 21st century skill - Frontiers in Education (2024). Peer-reviewed. [Reliability: High] ↩︎ ↩︎2

  14. The Knowledge-Creating Company - Nonaka, I. & Takeuchi, H. Oxford University Press (1995). Foundational work in organizational knowledge creation theory. [Reliability: High] ↩︎ ↩︎2

  15. The Tacit Dimension - Polanyi, M. University of Chicago Press (1966). Foundational work in tacit knowledge research. [Reliability: High] ↩︎

Psychology
Psychology AI Education Learning Metacognition protégé-effect tacit-knowledge prompt-engineering
This post is licensed under CC BY 4.0 by the author.