The best use of AI in the classroom isn't answering questions — it's building practice grounds

Education 2026-08-12 · Satsuma Creative · 9 min read

Three topics, three types of lessons — and only after finishing did I realize they share the exact same skeleton: AI lays out the options, humans make the call. AI's greatest value in the classroom isn't answering questions; it's turning "let the students rehearse it once" from an idea too expensive to attempt into something you can do in a single class period.

I wrote three pieces for the education section of Mandarin Daily News about how generative AI can enter the classroom. Three completely different topics: rewriting a textbook passage, planning a neighborhood walking tour, and teaching safety education.

It wasn't until I was wrapping up the third piece that I realized all three were really the same article.

Different subjects, different grade levels, different learning objectives — but the underlying skeleton was identical. And that skeleton is the same one I use when helping companies adopt AI.

This piece puts all three settings side by side and explains what that shared skeleton is. The originals ran in Mandarin Daily News; links are at the end.

Three settings

Setting one: one passage, three ways to play it

Every teacher has materials on the shelf they've used for years. There's nothing wrong with them — it's just that sometimes you wish the same content could be experienced a different way.

Doing that used to mean designing from scratch: writing worksheets, building scenarios, drafting discussion questions. All of it takes time, and time is exactly what teachers don't have.

One narrative passage with characters and a plot can be turned into three different formats by AI:

Rewrite it as a three-to-five-minute short play with narration and dialogue, performed in groups. The kids no longer just read the story — they *become* the people in it, voicing a character's hesitation themselves.

Or design three puzzles of increasing difficulty, with every answer hidden in the text. Reading shifts from "because I was told to" to "because I need to solve this" — the kids flip back to the textbook on their own.

Or write ten quick-response quiz questions at two difficulty levels. The last ten minutes of class, usually spent copying notes, becomes a small competition.

One passage, three completely different classrooms. What the teacher spends isn't several evenings of re-planning — it's a few minutes of waiting and choosing.

Setting two: one walking tour, four sticking points

Walking tours have become popular in elementary schools in recent years, but designing one raises a few familiar sticking points: I'm not a local history expert — what's actually here worth teaching? Once I know, how do I string scattered spots into a route that means something? On site, how do I get kids to observe closely instead of just strolling past? And afterward, how do I gather all those scattered discoveries into a reflection worth having?

Those four sticking points map neatly onto four things AI can help with: inventory, sequencing, task design, and pulling the discussion together.

Inventory means listing the neighborhood's features across three categories — nature, history, daily life. But that's a menu of options, not a decision about what to teach. The teacher looks at the list and picks: the park and the old trees resonate most with third graders; the market and the historical planning are too abstract, so leave them out for now.

AI lays out every possibility; the teacher chooses. That choice is something AI can't give you.

Sequencing means connecting the chosen spots into a route with a theme, not just a walk. One time it proposed a through-line called "Who gave the sky back to Minsheng Community," and the core insight it caught was this: the comfort of this place didn't happen naturally — a group of adults deliberately designed it sixty years ago. The order ran from large to small, far to near: the park, the street trees, then look up at the sky.

The *meaning* of the through-line is set by the teacher; AI just turns that meaning into one concrete stop after another.

Setting three: let students actually get scammed once

Information security gets taught every year, but when students actually encounter it, they may still click that link.

The reason isn't hard to see: the message is "don't click," but what kids are missing is the felt experience of being pushed along.

So the third piece is about this: using AI to build a practice ground where falling down doesn't hurt.

The raw material is a message anyone might get on their phone: "Congratulations, you've won! Click the link to claim your new phone — offer expires in three hours!"

Layer one: ask AI to write a 300-word short story based on that message. The protagonist is a fifth grader; it starts the moment the message arrives, captures their hesitation and excitement, follows them step by step as they do what they're told, and only at the end do they realize they've been scammed. The key instruction is "no moralizing; end at the moment of realization" — leave the conclusion for the students to draw. After reading, one question is enough: at which step did he get hooked? The debate that follows beats any lecture.

Layer two: have AI play the person sending the message while students play themselves, going through a real exchange. That's when students discover the scammer doesn't ask for information right away: first the congratulations, then "we just need to verify your identity." Every time they're questioned, the scammer backs off a step and offers another perfectly reasonable explanation.That feeling of being nudged along can't be put on a slide. The moment they catch on, have AI explain which tactic it just used — "manufacturing urgency" or "impersonating an official body." An instant debrief.

Layer three: build it into a five-level web mini-game. Each level is one scenario with three choices; after choosing, players learn what happens and why, with a score at the end. What separates it from an ordinary multiple-choice quiz: a wrong answer doesn't just get an X — you see what happened next.

The shared skeleton

Put the three topics side by side and the process is identical:

1. Give it material you already have. A textbook passage, a neighborhood, a scam message. You're not brainstorming from zero — you're using what's already in hand.

2. Ask for options, not answers. Three ways to play it, three categories of resources, three layers of the game. What you want is a list to choose from.

3. A human chooses. This step can't be outsourced. More on why below.

4. Turn it into something students have to do. Performing, solving puzzles, quick-response rounds, observation tasks, talking to a scammer, scoring through levels. What they share: students have to move, not just sit and listen.

5. Pull it together into a discussion. Gather the scattered discoveries into a reflection. For the walking tour, that's "why is the place I live the way it is." For scam awareness, it's lining up every hook the class found and picking the one you'd personally fall for.

Notice it — across these five steps, AI appears only between step one and step two. What it does is take over the time-consuming chore of generating options.

Why a "practice ground," not an "answer machine"

When most people think about AI in the classroom, the first thought is "write my test" or "write my lesson plan." That's treating AI as a faster printer.

But none of the three articles is doing that. They're turning static material into experiences students have to participate in.

And the reason few people did this before isn't that teachers didn't know it works — interactive teaching has always been on the wish list — it's thatthe production cost was too high. A role-play script, a set of puzzle clues, a mini-game that keeps score: each of those used to be several evenings of work. Fine as a one-off, impossible as a routine.

Once AI drives that cost down, what gets unlocked isn't "faster question-writing." It's that "let the students rehearse it once" goes from a luxury to something you can do in a single class period.

This is the point across all three pieces I think deserves the most attention:AI's value in education isn't in how much it knows — it's in making practice cheap.

Why that one step can't be outsourced

All three articles include a passage on the teacher holding the judgment. That's not a polite disclaimer — it's because AI really does get things wrong, and it gets them wrong in predictable ways.

It misjudges tone — a safety-education story that's too frightening, or the reverse: scam tactics described in such detail it becomes a tutorial.

It misjudges difficulty — a puzzle too hard or too easy for this age group. It can't know, because it has never met the thirty kids in your class.

It marks the wrong spot — on the walking tour, AI put stop two outside a shop on Fujin Street, but the real teaching point was "the old camphor trees along the way," not that shop. Leading the group in person, a teacher picks a stretch with room for everyone to stand and some shade.

These three errors share one trait:none of them make the output look wrong. The story reads smoothly, the puzzle looks legitimate, the map is neatly labeled. Catching the problem takes someone who knows what the real setting looks like going back over it.

This conclusion is identical inside companies

Only after finishing all three did I realize this is exactly what I tell businesses.

Swap "teacher" for "account manager," "textbook passage" for "quote," "students" for "clients," and the whole process holds unchanged: AI lays out the options, humans make the call, and the output has to be signed off by a person.

And the most expensive mistake is the same on both sides —it's not that AI says something wrong; it's that AI says it smoothly and nobody checks.

In the classroom, that's a script with a character written off-key — the kids perform it, and only afterward does the teacher realize the values it conveyed were off. In a company, it's an analysis with beautiful numbers or an email with perfect tone, sent before anyone notices the premise was wrong.

So whether I'm working with teachers or with companies, the last segment of the course is always the same: not more prompts, but a clear account of which judgments you can never hand over.

All three original pieces ran in the education section of Mandarin Daily News:

The scam-awareness game from the third piece is playable here:satsumacreative.tw/aq/demo0720.html. The quick-response mini-game from the first piece ishere. Both were produced by the very prompts described in the articles — no additional coding.

If you're a teacher or you plan professional development for a school, this material comes in half-day and full-day versions; the corporate training version uses the same skeleton, applied to your own workflows.Course details here。