The Scaffolded AI Tutor: Why AI That Won't Give Answers Belongs in the Classroom

After lunch on Tuesday, a P3 General Studies teacher opens the monitoring panel at the back of the room. One student has pasted the whole prompt "Why should we save water?" and three seconds later a complete answer fills the screen. She hits pause, walks over and asks: "After reading that, can you tell me one example of saving water at home in your own words?" The student freezes — what he learned was not the concept but how to paste. Another student on the scaffolded tutor gets only: "Does the question want causes or examples? Which have you thought of so far?"

For student-facing AI, the first design question is not how accurate the answers are but how to avoid ghostwriting. A scaffolded tutor turns support into adjustable steps; the teacher sets how high the steps go, and the student still does the thinking at every stage.

The cost of instant-answer AI in class

Paste the question, get the full answer — the student learns the interface, not the subject. Research and classroom observation agree: instant complete solutions weaken retention and transfer because cognitive load is outsourced and the brain never struggles or corrects. This is not a ban on technology; it is recognising that an AI tutor without scaffolding is essentially a ghostwriting tool.

Scaffolding: narrowing the spotlight, not switching it off

Scaffolding in teaching is like leading a hike: the guide does not carry you uphill, but points to the next marker at a fork, hands you a pole on a steep section, and if needed shows how to place one foot — the rest of the path is still yours.

Edor.ai's scaffolded tutor uses three adjustable gates, from light to heavier support:

Signpost (clarify) — Restate what the question asks; confirm what the student already knows and where they are stuck. No answer, only helping them see what is being asked.

Handrail (point) — Indicate which concept or method comes next, without full calculation or sentence ghostwriting. Like saying "this links to variable control from last week", not writing out the experiment steps.

One-step demo (partial) — Do one small piece and leave the critical step to the student. Write the first line of working, or offer a word frame for an opening sentence, then stop with "you try the next bit".

Full demonstration only when the teacher explicitly widens the task — usually revision or review, not first practice before an assessment.

Good scaffolding is not less help; it is help exactly where the student is one step from finishing alone.

Teachers set the support ceiling

Every "tutor task" needs a support ceiling: practice before an assessment might stop at handrail; revision week might allow one-step demo or full demonstration. The ceiling is set by the teacher by class, subject and phase — not one platform default.

The task card states clearly: topic scope, deadline, support ceiling, whether school notes may be referenced. Students who know the rules probe less often for direct answers.

Dialogue examples

P3 General Studies: saving water

Student: Write my answer for "Why should we save water?"

Tutor (signpost): The question asks for reasons. Can you name one way water gets wasted at home?

Student: Long showers.

Tutor (handrail): Right — long showers waste water. The question also wants "why we should save" — besides waste, what other consequences?

Student: Bills go up, and water might run out?

Tutor (one-step demo): Good. You could start with "Because… we should…" — Because long showers waste water, so we should — finish that sentence?

The paragraph the student submits is their own; every tutor line is a question or partial frame.

Form 1 science: controlled variables

Student: Which is the controlled variable in this experiment? Just tell me.

Tutor (signpost): The question lists three variables: sunlight, seed type, amount of water. To find "controlled variables", first ask: which factor is the scientist testing?

Student: Sunlight.

Tutor (handrail): If the scientist is testing sunlight, what should happen to the other two factors?

Student: Keep them the same… so seed type and water amount stay the same?

Tutor: Right. The controlled variables are the ones held constant. Can you name both?

Essential teacher-side features

Live monitoring — See the whole class's conversations; pause individual students (as in the opening scene) before ghostwriting spreads across the room.

Flag and follow up — Mark conversations that need attention and handle them at break or after school, rather than discovering the problem only from submitted work.

Misconception summaries — The platform collates class-wide sticking points (e.g. treating an irrelevant variable as controlled) for targeted teaching next lesson — more efficient than reading every chat.

Safety floor

Student conversations pass an input guard (inappropriate requests, personal data), a student-mode system prompt and output review; suspicious content is masked and flagged for administrators. The tutor must not request or output personal data, or help circumvent assessment integrity rules.

Common misreadings

• The claim: Not giving answers makes AI useless; students will use a search engine instead. What actually happens: Search engines also answer instantly. A scaffolded tutor's value is questioning aligned to your teaching sequence, with a monitorable conversation trail so teachers see where misconceptions sit.

• The claim: The higher the support ceiling, the happier students are. What actually happens: Full demonstration before an assessment is ghostwriting. Ceilings should follow the teaching phase — wider in revision, tighter on first attempts.

• The claim: Writing "do not give the answer" in the prompt is enough. What actually happens: Generic models default to answer mode. You need platform-level scaffolding, support gates and teacher ceilings; a single prompt line is unreliable.

Three things to try on Monday

1. Set a support ceiling for your next tutor task (e.g. handrail only) and explain the three gates to students in the task brief.
2. Open the monitoring panel for ten minutes in class and note the phrases students use to probe for direct answers; remind them orally next lesson.
3. Review last week's misconception report, pick one class-wide sticking point and spend five minutes clarifying it next lesson — let AI conversation data feed teaching, not just sit in archives.