The Maps We Copied — and what they teach us about AI in education

Remember when the geography teacher assigned us to draw a map of the country's mountain ranges? And a map of the rivers of Africa and South America?

Where did we get the information — and the maps themselves? From books, or later from the Internet. We literally copied the maps from there onto our paper.

And the teacher didn't accuse us of being unoriginal. The teacher never expected us to create those maps from scratch. More than that: the teacher didn't actually care about the map, and had no expectation that we would become cartographers or visual artists.

What the teacher made us do was interact with the object of study. Not the map itself — the geography of the country or the continent: the tallest peaks, the longest rivers, where things sit in relation to each other. Through a mechanical, physical process of reproduction and engagement, we learned the object of study. The map was the medium. The geography was the lesson.

Piaget would have recognized the move: we learn by constructing our own version of the world, not by receiving one. And Bruner named the ladder we were climbing without knowing it — from doing, to image, to symbol: enactive, iconic, symbolic. The hand traced, the map took form, and the geography became something we could finally say.

Learning research would later map the terrain around this. Physically tracing a representation eases the cognitive load and anchors attention near the hand — a practice as old as Montessori's sandpaper letters. Making a representation you generate — selecting, organizing, placing — is one of the better-documented ways to learn anything. And nobody, as far as I can find, ever ran the controlled study on copying the mountain ranges out of a textbook. Our teachers didn't know the mechanism either. What they knew was how to design an assignment in which producing the thing forced us to keep looking at the thing. And the copied-versus-generated distinction is exactly where AI gets interesting — because the AI version of this assignment is generative by construction.

AI in education can be designed to play exactly the role that map assignment played: a tool for interacting with the object of study in fun and creative ways, so that learners learn it in the process. Not is — can be. The distance between those two verbs is the whole craft of assignment design. Seymour Papert called the underlying idea constructionism — learning by making something meaningful — decades before anyone prompted an image model. (Papert's own record carries a useful caution: making did not automatically transfer into general thinking skills. The learning lives in the specific engagement — which is precisely the claim here.)

Three decades past the deadline

Here is the uncomfortable part. AltaVista opened the full text of the Web to anyone in December 1995 — three decades ago. And let's be precise about what it changed. It did not make remembering obsolete: retrieving knowledge from your own memory remains one of the best-evidenced ways to make learning stick. What it made hard to justify is a different kind of schoolwork: the assignment whose entire intellectual demand is finding a fact a search engine surfaces in seconds — what, where, when, who — and copying it into a deliverable.

So why do we insist on assignments built on rudimentary retrieval and reproduction of information — the lowest rung of every taxonomy of learning we have, from Bloom's remember to Webb's recall and reproduce? The tools made those tasks trivial thirty years ago. The assignments survived the tools.

Educators saw it at once: the very year AltaVista launched, Bernie Dodge and Tom March built the WebQuest around one rule — a Web assignment must go beyond fact-finding. The redesign has been available for thirty years. Mass schooling mostly didn't take it. (When TNTP followed nearly 4,000 students in 2018, it found over 500 hours a school year going to assignments below grade level — practitioner data, not peer review, but that is the size of the habit.)

The same assignment, with AI in the room

But before we get too technical, let's go back to our maps and bring AI into the picture. What if the teacher, instead of having us draw the map, assigns:

1. Before you generate anything, identify the ten highest peaks from authoritative geographic sources and sketch where you expect each to sit. Then create the map with AI — and audit every peak for name, rank, altitude and location against your sources, recording each error, until the map satisfies the evidence (not the model).
2. Draw a line connecting the highest peak of each range in descending order of altitude — starting at the highest, ending at the lowest.
3. Create an infographic comparing our mountain ranges and peaks with the Andes, the Alps, and the Himalayas — every figure in it verified against a source you can name.
4. Produce a selfie of yourself on top of each of the ten highest peaks — audited for the real mountain: the right vegetation, the right snow line (if any), the right air, the right view behind you. Every plausible-but-wrong detail you catch is geography learned.
5. Finally, with the AI off: on a blank outline map, place the five highest peaks from memory and explain the pattern you now see.

It's not hard to imagine that, as of 2026, most commonly available AI tools will choke on these — and fail to produce the expected results unless prompted carefully and repeatedly until the output is right.

No problem with that. That's the assignment working.

Because to get peak number seven in the right place at the right altitude, the learner has to know where peak number seven is — and catch the AI when it puts it somewhere else. To make the selfie environmentally correct, the learner has to find out what the top of that mountain actually looks like. Every iteration, every correction, every "no, that's wrong, fix it" is another pass through the object of study. Notice who is doing the checking, the comparing, the insisting: not the tool — the learner.

The strongest field experiment we have says the same thing from the other direction. High-schoolers with unrestricted ChatGPT scored 48% better during practice — and 17% worse than peers who never had AI, once the AI was taken away at exam time. The same AI, redesigned to give hints instead of answers, erased the harm (Bastani et al., 2025). So the iteration can become the engagement — if every round requires the learner to inspect the subject, detect what is wrong, decide what changes, verify the fix against something outside the model, and explain why the new version is better. Design it any looser, and the tool quietly absorbs the very thinking the learner was supposed to practice. The map assignment always knew this. The effort was never in the copying — it was in the engagement, and the engagement is what we keep.

None of this makes the teacher smaller. The teacher's real work was always the design of the zone — Vygotsky's term — where the task sits just beyond what the learner can do alone, held up by scaffolding (Wood, Bruner and Ross's word) until it can come down. An AI assignment is zone design too. And iterating against verification is nothing more exotic than deliberate practice in new clothes: effortful repetition at the edge of ability, with feedback that doesn't flatter.

And notice something else: nobody in this assignment is accused of cheating for using AI, just as nobody accused us of copying the rivers of Africa. The tool is declared, the tool is the medium — and the learning is designed to happen anyway, on purpose.

We didn't become cartographers. We learned where the mountains were. Our learners won't become prompt engineers. They'll learn what they were actually studying — if we design the assignment so the AI is the map, and never the lesson.

Grade the journey, not the map

If the artifact no longer proves learning, what does? Four layers. The artifact — what the human-AI system produced. The process — what the learner decided, rejected and verified along the way. The reflection — what the hardest error taught. And transfer — what the learner can now do with the AI turned off. This is not hypothetical: it is how our Smoother methodology evaluates learning, and how the aiLearning Challenge evaluates it as well — process weighted over product, and a live defense with no AI in the room as the final gate.

None of this is orphan practice. The designer's shelf is full — Gardner's multiple intelligences, read properly as a reminder to open many doors into the same object of study and never as a machine for labeling learners; design thinking and design sprints; agile's iterative prototyping; project- and problem-based learning, which Stanford's PBL Lab frames as five P's: problem, project, product, process, people; and the instructional-design canon that professional educators build with — Gagné's conditions of learning, Kolb's experiential cycle, Merrill's first principles, van Merriënboer's 4C/ID, ADDIE, SAM, Understanding by Design. These are the educator's tools, not the learner's ceremony: they do their work on the designer's side of the assignment, which is exactly where Smoother puts them.

The object of study was never the map. It still isn't.

This essay is my parallelism and professional position, stated as mine. The dates, the taxonomy frame, the drawing-and-tracing distinction and the Bastani result are verified against the sources below; the extension of constructionism to AI-mediated assignments is my expert opinion, not a research finding. — Carlos Miranda Levy

Four perspectives

References

Cited in the essay:

• Anderson, L. W., & Krathwohl, D. R. (Eds.) (2001). A Taxonomy for Learning, Teaching, and Assessing. Longman.
• Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122.
• Bruner, J. S. (1966). Toward a Theory of Instruction. Harvard University Press.
• Cromley, J. G., Du, Y., & Dane, A. P. (2020). Drawing-to-learn. Journal of Science Education and Technology, 29(2), 216–229.
• Dodge, B. (1995). Some Thoughts About WebQuests. San Diego State University.
• Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406.
• Ginns, P., Hu, F.-T., Byrne, E., & Bobis, J. (2016). Learning by tracing worked examples. Applied Cognitive Psychology, 30(2), 160–169.
• Papert, S. (1980). Mindstorms. Basic Books; Papert, S., & Harel, I. (1991). Situating constructionism. In Constructionism. Ablex.
• Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning. Psychological Science, 17(3), 249–255.
• TNTP (2018). The Opportunity Myth. (Practitioner research, not peer-reviewed.)
• Van Meter, P., & Garner, J. (2005). The promise and practice of learner-generated drawing. Educational Psychology Review, 17(4), 285–325.
• Vygotsky, L. S. (1978). Mind in Society. Harvard University Press.
• Wammes, J. D., Meade, M. E., & Fernandes, M. A. (2016). The drawing effect. Quarterly Journal of Experimental Psychology, 69(9), 1752–1776.
• Webb, N. L. (2002). Depth-of-Knowledge Levels for Four Content Areas; Webb, N. L. (1997). Research Monograph No. 6. CCSSO/NISE.
• Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89–100.

Further reading — including the OECD 2026 Digital Education Outlook, UNESCO's generative-AI guidance, the Chen & Cheung 2025 meta-analysis, Olney & Cade on learning by correcting AI errors, the Logo transfer literature (Pea & Kurland; Swan; Scherer et al.), and van Merriënboer's 4C/ID — is collected in the essay's evidence file.