The Final Stage of Learning Is Teaching

Marielyn Wong
6 days ago
9 min read

Now Students Can Teach the Machine

Editor's note: This is a long read — around 3,500 words. It started as a LinkedIn post, became an essay, and ended up as the clearest articulation of why we're building what we're building at Noodle Factory. We've formatted it for the blog, but haven't shortened it. Some arguments need the space they need.

For most of the last few years, the question we have asked of artificial intelligence in education has been: how do we use these systems to teach students?

It is a good question. Research has filled knowledge bases in an attempt to answer it. At Noodle Factory, it is the question my co-founder and I have asked for years as we built the platform.

But increasingly, that question may not be the most interesting one.

A more interesting question — now that the tools are mature enough to support it — is the inverse. What happens when students use AI not to be tutored, but to tutor? What happens when a student is offered the opportunity to not consume an AI tutor, but to build one?

That shift is significant. The risk is real too. But I want to make the case that we ought to be doing more of this, not less — provided we do it well.

Why teaching is the final stage of learning

The pedagogical claim underneath all of this is much older than AI: we learn more deeply when we teach.

The cleanest articulation of this idea comes from medical education, where William Stewart Halsted's late-nineteenth-century apprenticeship model was eventually compressed into the slogan "see one, do one, teach one" (SODOTO). The slogan endured because it captured something true. Procedural mastery cannot be acquired by passive observation, nor by repeated practice alone. What completes it is the act of transmitting it to the next learner.

A participant in one of my own professional development workshops once pressed me on a topic I had taught for years. On the third attempt at explaining it, I realised something I had not realised before: I had taught it without ever truly understanding one of its more nuanced aspects. The workshop had gone fine. I had prepared, organised, tailored, and presented. But fielding a real question in real time exposed a gap that my preparation had hidden.

Every teacher knows a version of this story.

It is worth being honest about SODOTO's limits, though. Medical education has largely moved away from it because the framing didn't survive contact with patient safety. The contemporary alternative — Entrustable Professional Activities — preserves the deep insight while protecting against naive misapplication. That is exactly the spirit in which to import this frame into a debate about AI in education. Not as a slogan to repeat; as a structure to refine.

The cognitive science literature on learning by teaching is broader than any single tradition. The protégé effect — the finding that we exert more cognitive effort and learn more reliably when we believe we are, or will be, responsible for teaching another — has been documented across decades of research (Fiorella & Mayer, 2016). The most-cited demonstration in the AI-adjacent literature, Vanderbilt University's "Betty's Brain," found that eighth-grade students who taught a digital agent made stronger learning gains than students in a control condition. A subsequent meta-analysis (Kobayashi, 2019) added a critical finding: the protégé effect requires the expectation of teaching to be set before the learning begins. Retrofitting a teaching task on top of a study task does not produce the same gains.

To these traditions, we can add generative learning theory (which lists "teaching" as one of eight strategies that produce durable learning by forcing the learner to actively construct meaning), Bloom's revised taxonomy (which places "Create" at the pinnacle of cognitive engagement), and Seymour Papert's constructionism (which argues that learning is most generative when the learner produces a public, shareable artefact).

Five theories. One mechanism. The reason teaching deepens learning is that teaching makes your own thinking visible to yourself, with sufficient force that you cannot quite ignore it.

What changes when the teaching is the building of an AI

The most useful way to think about a student-built AI tutor is as two distinct learning loops.

The first is the building loop. Configuring an AI tutor — selecting content to include, defining learning outcomes, structuring a pathway, anticipating where a learner will get stuck — is a teaching method in everything but name. The student is the explainer. They are confronting their own knowledge gaps in the process of organising knowledge for someone else. The cognitive demands are higher than in the original teachable-agent studies, because the system being taught is generative — its responses must be anticipated and guided, not just selected from a menu.

The second is the using loop. When a student subsequently engages with the tutor they built — using it to revise and test their own understanding — they are now being taught, in a sense, by their own externalised mental model. They see their own organisation of the material from the outside. The artefact they built last week becomes the teaching object they encounter this week.

These are not alternative descriptions of the same activity. They are distinct cognitive episodes arising from a single creative act. That dual-loop architecture is, I would argue, a genuinely novel contribution of the student-as-creator model.

Three structural features make it different from prior "learning by teaching" approaches: scalability (a student who builds a tutor creates a learning environment that can serve many future learners, not just one classmate), iterability (a study guide is finished when it is finished; an AI tutor is not — each interaction is a feedback signal), and personal investment (a learner who has designed a tutor, named it, and watched it interact with peers has a different relationship to the content than one who passively studied it).

The mirror tutor

I want to introduce a name for what a student-built AI tutor is at its best, because I think the absence of a name has prevented us from seeing it clearly.

I am calling it a mirror tutor.

A mirror does not tell you what to think; it shows you what you already are. A mirror tutor, in the same spirit, is an artefact that reflects the student's own knowledge structure back to them — including its biases, omissions, and structural assumptions — with sufficient clarity that they cannot quite ignore what they see. It is not an oracle. It is not a substitute for instruction. It is a surface against which the student's own understanding becomes visible.

Three properties distinguish a mirror tutor from prior teachable agents.

The first is authorship. In the original Vanderbilt studies, the student "taught" Betty by selecting from pre-supplied concepts; the agent's structure was largely given. In a mirror tutor, the structure itself is the student's work. Whatever depth their understanding has, the tutor will inherit; whatever shallowness, the tutor will mirror that too. The mirror is honest in a way a multiple-choice exercise is not.

The second is intimacy. A mirror tutor might live in the place where a student already does most of their learning: on their phone, in the late evening, between commitments. The relationship is closer to a reflexive journal than to a worksheet.

The third is reflexive feedback. A mirror tutor, when it fails to teach you, gives you feedback about your own understanding — because you are simultaneously the author and the consumer. You cannot blame the explanation, because the explanation is yours. You cannot blame the curriculum, because you set the curriculum. That reflexive loop is uncomfortable. It is also, I suspect, where the deepest learning happens.

Rozenblit and Keil's (2002) work on the illusion of explanatory depth — the well-replicated finding that we believe we understand things we cannot actually explain in detail — is the precise condition the mirror tutor is built to expose. Ask a student to summarise a concept, and they will tell you they understand it. Ask them to build a tutor that teaches it, and the illusion will not survive the construction.

Three honest objections

1. Equity

The students best positioned to build a good AI tutor are, on average, those who already arrive with stronger metacognitive habits, more comfort with technology, and richer prior preparation. These are also the students more likely to afford paid AI access if their institution doesn't provide it. If we deploy student-built tutors as enrichment for those already enriched, we widen the gaps the field claims to want to close.

There is a partial counter-current: building externalises the gaps that lecture-based instruction hides. A student whose understanding is shallow will encounter the limits of their own knowledge earlier in the process, while there is still time to intervene. But this is only a corrective force if the institution responds to it as one rather than as a sorting mechanism.

The student-as-creator model must be designed for equity, not assumed to be equity-neutral. Formative assessments before the build. Scaffolding that reduces cognitive barriers. Peer review that distributes learning rather than just ranking outputs. None of this is invented for AI; all of it is standard equity-conscious pedagogy. What is new is that the artefact being built is more powerful and more public than anything the existing literature has had to account for.

Educational technology providers do not get to claim equity as a feature unless they build with the constraints of equity fully in mind.

2. Cognitive load

John Sweller's cognitive load theory is right within its domain: working memory is limited, and asking novices to simultaneously master content and design instruction is a recipe for overload. Asking a first-year student to build an AI tutor on a concept they encountered last week is, in cognitive load terms, almost guaranteed to fail.

The objection bites less when the building task is sequenced after a defined threshold of knowledge acquisition. The student-as-creator model is best understood not as a replacement for direct instruction but as the activity that follows it — once the student has a base of knowledge stable enough to be elaborated on rather than assembled in real time.

Sequence matters. The same activity is a cognitive overload at one stage of a course and a cognitive corrective at another.

3. Disciplinary boundaries

The "see one, do one, teach one" frame emerged from procedural training. Importing it into other fields requires care.

For procedural disciplines — nursing, accounting, engineering — the model maps cleanly. The tutor encodes a procedure; correctness is largely a matter of consensus; success is measurable.

For conceptual disciplines — philosophy, history, literary interpretation — the model operates differently, and more generatively. The act of building a tutor in these disciplines requires the student to make and defend an interpretive stance. The artefact reveals not only what the student knows, but how they read. That is, to my mind, the single most pedagogically interesting move that student-built tutors enable in the humanities.

For contested-knowledge disciplines — ethics, public health, political science — building a tutor is a values exercise as well as an epistemic one. Faculty review is not optional; it is the constraint that protects the field from being misrepresented to the next cohort. The bar is higher. That makes it more interesting, not less worth pursuing.

Five questions before you begin

If you are an educator or L&D professional considering this approach, these are worth asking before you assign students to build AI tutors:

Have you declared the teaching intent up front? The protégé effect requires this expectation to be set before the learning begins — not retrofitted.
Have you set a knowledge threshold before the build begins? Requisite formative assessment is not gatekeeping; it is metacognitive priming.
Have you separated assessment of the artefact from assessment of the learning? A polished tutor can hide a shallow learner; a rough one can reveal a deep one.
Have you designed for iteration, not one-shot delivery? Tutors that are expected to evolve are tutors whose imperfections are part of the learning.
Have you made expert review a feature, not a bottleneck? Faculty as quality assurers, not line editors. Peer calibration before deployment.

What this means for Noodle Factory

For the first several years of the platform, our five pillars — structured pathways, AI-powered tutoring, learner personalisation, institutional governance, and a shareable ecosystem — were operationalised in one direction. Educators authored. Students consumed.

The student-as-creator move is not a sixth pillar. It is the inversion of the existing five. Pathways become student-authored. Tutoring becomes student-configured. Personalisation becomes self-personalisation. The ecosystem moves from a system of distribution to a system of co-creation.

We have chosen to support this inversion. Time will tell us whether it was the right bet.

A closing invitation

The most important AI literacy we can teach over the rest of this decade may not be how to use AI — but how to build something that holds up under scrutiny, including your own.

A mirror tutor is a small example of a larger principle: the artefacts students build are the reflective surfaces against which their understanding becomes clearer and more legible, to themselves and to us.

As in the original Socratic academy, the final stage of learning is, and always was, a form of teaching. We now have the chance, for the first time, to test this at scale.

I am looking for collaborators — in higher education and workplace learning — to design and run an action research programme to test the hypotheses raised here. If you would like to be involved, I would like to hear from you: jim@noodlefactory.ai

Dr Jim Wagstaff is a co-founder of Noodle Factory, an AI-powered teaching and learning platform, and a researcher in higher education and digital pedagogy. He writes from Singapore.

Suggested citation: Wagstaff, J. (2026). The Final Stage of Learning Is Teaching: Now Students Can Teach the Machine. Self-published practitioner essay. Singapore.

This essay also appears as a LinkedIn Article on Jim Wagstaff's profile.