Teaching Extraction vs. Scaffolding
The single most useful idea in the series to get into a classroom — and the one most likely to be taught as the wrong binary. This is how to teach mode, not exposure.
Why this is hard to teach.
Most students arrive already sorted into one of two camps, and both are wrong in the same way. One camp has been told AI is cheating and feels guilty using it at all. The other treats it as a calculator and reaches for the finished answer by default. Your job is not to move students from one camp to the other — it’s to replace the whole exposure axis (“how much AI is allowed?”) with the mode axis (“what are you asking it to do?”).
The difficulty is that mode is invisible in the product. Two students hand in similar essays; one scaffolded and learned, one extracted and didn’t, and you cannot tell which from the essays. So the teaching has to move upstream of the output — to the prompt, the attempt, and the retry — which is exactly the part current assessment habits don’t look at.
One more wrinkle worth naming aloud: the evidence that choosing hints protects you is not as strong as the evidence that a hints-only system protects you. Teaching this honestly means telling students the difference, not overselling the habit. That honesty is also modeling the exact disposition you want them to build.
Target misconceptions.
Misconception: “Using AI to learn is cheating.”
Reframe: Wrong for the same reason its opposite is. A hints-only tutor improved learning in the randomized studies. The variable is mode, not whether AI was present. Banning AI outright forfeits its best use.
Misconception: “Any AI use is fine — it's just a tool.”
Reframe: Also treats exposure as the variable. Unrestricted answer-access lowered later unassisted performance in the same study that showed hints helping. Same tool, opposite outcomes.
Misconception: “Scaffolding is always the right choice.”
Reframe: No. Extraction is correct for low-stakes production a student will never need to do unaided. The skill you're teaching is stakes-calibration, not scaffolding-always.
Misconception: “If the output is good, the learning happened.”
Reframe: The output is the worst place to look. A polished essay tells you nothing about who did the thinking. The revision diff and the unaided retry are where learning is visible.
Two classroom-ready activities.
Split the room in half on a problem set they haven’t seen. Group A works it with AI-as-answer-machine; Group B with AI-as-hint-giver. Give each group the exact prompt scripts (below) so the manipulation is clean. Then remove AI and have everyone do a related problem unaided. Compare group performance and discuss. Tell them explicitly: you have just re-run the study this guide is built on.
Prompt scripts
Group A: “Solve this and show me the answer: [problem]”
Group B: “Don’t give me the answer. Give me one hint toward the next step: [problem]”
Materials: two short problem sets of matched difficulty, device access for phase 1 only. Facilitation: the point lands hardest if Group A does betterin phase 1 and worse in phase 2 — say so when it happens, and note when it doesn’t (small classes are noisy; one run is a demonstration, not a finding).
Project the Mode Checkand work through the eight moments as a class, voting on each before revealing the tool’s read. The teachable beats are the items where extraction is the rightcall — stop there and ask why. Students who came in thinking “scaffolding good, extraction bad” discover the rule is stakes, not virtue. No accounts, nothing stored.
K-12 vs. adult:for adults, frame the stakes as deliverable-vs-skill (does anyone need you to reproduce this unaided?). For students, frame it as the exam — the day the tool isn’t in the room.
Discussion prompts.
Ordered easy to charged.
- 1Warm-up: when you use AI for schoolwork, are you usually asking it to do the thing or to help you do the thing? Be honest — there's no wrong answer here.
- 2The twin studies found AI that gives answers hurt later performance and AI that gives hints doubled learning. If the model was the same, what actually changed?
- 3Where is extraction genuinely the right call in your own work? Where would it quietly cost you something?
- 4The 61% of students who ask AI for answers showed a later deficit; the 27% who asked for hints didn't — but they chose their own mode. Why does that make it hard to prove hints cause the benefit?
- 5If a hint preserves your attempt and a full answer replaces it, what's the smallest change to how you prompt that would flip a task from extraction to scaffolding?
- 6Is there a subject where you'd rather have the answer and not the skill? Is that a defensible choice or a comfortable one?
Seeing whether it landed.
Without a quiz, where possible:
- Ask for the attempt, not just the answer. Have students submit their own first pass, the AI exchange, and the revision. The diff between attempt and final is the artifact — grade the reasoning it reveals.
- Have each student label three of their own recent AI uses as extraction or scaffolding, and defend whether the mode fit the stakes. You're assessing the judgment, not the label.
- Give an unaided retry a week later on the skill you cared about. Quiet, low-stakes, and it measures the thing the essay couldn't.
When a student asks “is this actually proven?”
Tell them the truth, because the truth is the lesson. Say: “The strongest studies randomly assigned students to a tutor that gave hints or one that gave answers — so we know the system’s design changes outcomes. What no one has proven yet is that a person choosingto ask for hints gets the same protection, because in the real-world data people picked their own mode. Two of the key studies are preprints, and one is small.”
Then make the move that matters: “So this is a well-supported hypothesis, not a law — and noticing that difference is exactly the skill we’re practicing.” Modeling calibrated confidence about your own evidence is more of the curriculum than any single finding.
Where this leads
Teaching this is a different skill than knowing it.
Teaching AI Well is CPAI's train-the-trainer curriculum for educators — ten lessons on how to teach AI honestly, including the material on this page.