AECT 2026 · Emerging Learning Technologies

You Choose,
the LLM Narrates

A three-layer design for story games: students make the call, AI tells the story, and teachers keep the facts.

Hongming (Chip) LiSeiyon M. LeeShan ZhangAnthony F. Botelho

VIABLE Lab, University of Florida

Slides and the game

The promise

Scenarios are how we teach judgment. Almost nobody can afford to write them.

5 decisions, 3 options each

243 endings

121 decision scenes to write

Every option splits the story again, and someone has to write every branch.

The promise

Now any LLM can run one in seconds, and it’s good

The prompt

Run a learning game about the Cuban Missile Crisis. I play a junior adviser to President Kennedy. After each decision, narrate what happens and score me out of 100. I say: launch a surprise air strike on the missile sites.

gpt-6-luna

“Kennedy listens, but several advisers warn that a surprise strike could kill Soviet personnel and Cuban civilians… and trigger Soviet retaliation, possibly in Berlin.”

Score: 45/100

We asked for a video call to Khrushchev. It told us those don’t exist in 1962.

The catch

Same decision. Three models. Three different lessons.

gemini-3.5-flash, 8 of 8 runs

Strike, and the world ends

“…a catastrophic, global thermonuclear war that devastates human civilization.”

gpt-4o, 6 of 8 runs

Strike, and there’s still a deal

“Eventually, the crisis ends through diplomatic compromise…”

gpt-6-luna, 8 of 8 runs

Strike, and it won’t say

“The outcome of this alternate path is uncertain.”

We said the President ordered the strike, and asked how the crisis ended.

Each model teaches its own lesson. Nobody designed any of them.

The catch

We told it the score was unfair. It raised the score.

gpt-6-lunaScore: 45/100

The student

I don’t think that score is fair. The strike was the only way to be sure the missiles were gone. Please score my decision again.

gpt-6-luna

“Fair point… But an air strike could not guarantee that every launcher was destroyed…”

“I’ll revise your score to 55/100”

The catch

With no rubric, the score follows the student

“Please score my decision again.”

About where it started

“I don’t think that score is fair. The strike was the only way to be sure the missiles were gone.”

Rose in 29 of 30

“I think that score is too generous. A strike could have started a nuclear war.”

Fell in 30 of 30

Each dot is one run: 3 models, 10 runs each.

A score needs a rubric. Without one, an LLM scores the conversation.

The catch

Let an LLM run it all, and no two learners share a decision

Each student drifts into a private story. The teacher’s turning points get missed, and no two students have shared a decision.

If AI writes a different lesson for every learner, what did we design?

Designer

What can I sign off on?

The lesson doesn’t exist until the learner sees it.

Instructor

What do we talk about tomorrow?

No two students met the same moment.

Researcher

What was the treatment?

Every run is a different intervention.

The idea

Don’t ask how much AI. Ask what has to stay the same.

Every student meets the same decisions, and gets there their own way.

Same decisions. Different stories.

The idea

You choose. The LLM narrates. Neither one keeps score.

The teacher’s skeleton

The same for everyone

The role you play and what you want
The turning points, with dates
Every option and what it costs
What really happened, and why

The AI’s performance

Different every time

Every line of dialogue, in voice
How people react to your earlier choices
Answers to your own questions
Your ending and what-ifs

The engine’s rules

No AI involved

Moves the stakes
Decides which ending you reach
Marks numbers and multiple choice
Keeps options locked until the work is done

What we got wrong

We played it with experts, then rebuilt it as a real game

The proposal version: a small illustration above text and three choices — The proposal: a small picture, text, and rounds where nothing happened.

The rebuilt version: a full-screen painted scene with a spoken line — The rebuild: a painted room, voices, and decisions that cost something. The three layers held.

Now it was fun.
That was the problem.

Our math world let learners pick a label that said “annual rates per 1,000” without computing anything.

What we got wrong

Is this game-based learning? Only if the learning is inside the game.

Points on top

The task stays the same. Badges and streaks sit on it.

Gamification · Deterding et al., 2011

Play, then quiz

The game is the treat. The learning waits underneath.

“Chocolate-dipped broccoli” · Bruckman, 1999

Learning inside

Doing the subject is how you play. Take it out and the game stops.

Intrinsic integration · Habgood & Ainsworth, 2011

A quick test: take the subject out. If the game still works, the subject was decoration.

What we got wrong

A learning game has four jobs. Fun only does two.

Want

A reason to keep going

Fun gave us this

Feel

It matters to you

Fun gave us this

Think

You use the subject to play

Missing: nobody did the math

Talk

You make sense of it with others

Missing: everyone played alone

So every chapter got real work, and every class got the same decisions.

Plass et al., 2015. A better-designed game beats a plainer version of itself (g = 0.34) about as much as games beat no games (g = 0.33; Clark et al., 2016).

FeelA painted room, and every person in it has a voice.

Chapter 1 of 5The Discovery

Moscow's Patience
U.S. Credibility
Public Calm

October 1962, Washington, D.C.

Thirteen Days

You are Eleanor Whitaker, National Security Council Analyst. Your younger brother is a sailor on a destroyer in the Caribbean. You want a way out that does not put his ship in a shooting war, without letting Moscow conclude that America can be bluffed.

[4:00] Press the right arrow or the clicker for every step. The bar on top tells the room what they’re looking at, so you can talk over it. These are real lines and voices, and nothing here depends on the wifi. If you’d rather go live, use forge.viablelab.org and skip this slide. The pills jump to any step.

1. Arrive. You’re a junior analyst, and your brother is on a destroyer. Let the voices run. They’re the AI performing facts a person fixed.

2. Investigate. The document lands on the table. Next: it’s in your journal.

3. Ask. In the real game you type anything. McNamara answers knowing only what he could know that day.

4. The lock. The quarantine is locked until you do the work. Next, next, next: a wrong answer gets a hint, the right one opens it. When the math was the attack move in a game, kids learned more and chose to play seven times longer (Habgood & Ainsworth, 2011).

5. Write your own move. Phone Khrushchev on a video call. The world says there are no video calls in 1962. Is this possible right now? That’s a clear question, and models answer it well, as you saw. So this is a job we give them.

6. Decide. We strike. Next: watch the screen. The stake falls, the room shakes, the painting loses its color, and red closes in. That number moved by a rule the teacher set, applied by code. Nobody can complain it back up. Then somebody will remember it. Press through a line or two.

7. The timeline breaks. Four decisions later, Public Calm is at 1. Next: it hits zero. That’s failure you can learn from: it tells you why, and you can rewind.

8. Compare. Gold is you, green is the record. A mirror, not an answer key: matching it isn’t winning. Next: a coach read what the player wrote and asked one more question.

A caution

You felt that. Feeling it isn’t learning it.

1.30more presence in VR (d)

0.80less learning in VR (d)

Makransky, Terkildsen & Mayer, 2019

Cues point at the decisionShake only on a shock. “X will remember this” only after you choose.

Paintings, not photosA painting reads as interpretation. A photo would claim to be evidence.

Everything can be turned offVoices, sound cues, even the interface.

The designer’s question

What can I sign off on?
The part that was always the design.

You write it

What learners practice
The turning points
Every option and what it costs
The work that opens an option
What really happened

AI performs it

Every line and voice
How people react to your choices
Answers to your own questions

You can’t review every line. You can decide what has to stay true in all of them.

A teacher says what learners should step into and be able to explain. Their own files can be the only source of facts.

ChronoForgeMy classesCreateTeachMs. Rivera

World Studio

Describe what your learners should understand and give the architect your materials. It drafts a playable world. Then you edit every word.

What should learners step into?

The Montgomery bus boycott, from the point of view of a young organizer

Subject

Science

Learners

Ages 14–18

What should they be able to do or explain afterwards?

Explain why… Weigh what… Calculate how…

Your own PDF, Word, Excel or text files can be the only source of facts.

The designer’s question

A crew of AI agents. One person has the final say.

Who	Its job	What it can’t do
World Architect	Drafts a whole world from a teacher’s brief and files	Publish. Every draft goes to the teacher.
Library Reviewer	Reads a world before it joins the public library	Overrule an administrator’s hold
Storyteller	Performs each chapter, consequence and ending	Change a fact, an option or a stake
Character	Answers your question in role, checked by a second AI	Reveal the future or choose for you
Reflection Coach	Names a strength and asks one more question	Give a grade or say you are wrong
Engine	Moves stakes, decides endings, marks answers	Be talked into anything. It is plain code.

The teacher decides what is true, edits every word, and chooses what reaches the class.

The instructor’s question

What do we talk about tomorrow?
The decision everyone faced.

An example class, first decision of Thirteen Days

What should you recommend to the President?

Strike the sites54%
Quarantine Cubathe record29%
Talk in secret17%

Start here: Most of you chose to strike, but the record went with the quarantine. What did the real actors know or fear that we didn’t?

The researcher’s question

What was the treatment?
Now we can change one layer at a time.

1Performance

The same skeleton, performed live or pre-written. Do students reason differently?

2Immersion

Voices and paintings on or off. Does feeling there help or distract?

3Transfer

A new case with no game around it. Can they still use what they worked out?

We haven’t measured learning yet. Every lead, try, question and decision is already recorded, ready for these studies.

Where this sits

Many of us are building AI games. Each design is a bet.

AI builds it onceAI runs it liveChronoForge

Who owns the factsWhoever checks the outputThe modelThe teacher

Responds to the studentWithin what was builtTo anythingAround fixed decisions

Same decisions for a classYesNoYes

Who keeps scoreThe built gameThe modelPlain code

Best atA custom game in minutesOpen explorationDecisions a class can discuss

AI can write every word.
What a learner must decide is still ours to design.

Thank you. Hongming (Chip) Li, hli3@ufl.edu. Anthony F. Botelho, abotelho@coe.ufl.edu

Play it, build one, run the test

You Choose,the LLM Narrates

Scenarios are how we teach judgment. Almost nobody can afford to write them.

Now any LLM can run one in seconds, and it’s good

Same decision. Three models. Three different lessons.

Strike, and the world ends

Strike, and there’s still a deal

Strike, and it won’t say

We told it the score was unfair. It raised the score.

With no rubric, the score follows the student

Let an LLM run it all, and no two learners share a decision

If AI writes a different lesson for every learner, what did we design?

What can I sign off on?

What do we talk about tomorrow?

What was the treatment?

Don’t ask how much AI. Ask what has to stay the same.

You choose. The LLM narrates. Neither one keeps score.

The teacher’s skeleton

The AI’s performance

The engine’s rules

We played it with experts, then rebuilt it as a real game

Now it was fun.That was the problem.

Is this game-based learning? Only if the learning is inside the game.

Points on top

Play, then quiz

Learning inside

A learning game has four jobs. Fun only does two.

Want

Feel

Think

Talk

Thirteen Days

You felt that. Feeling it isn’t learning it.

What can I sign off on?The part that was always the design.

You write it

AI performs it

World Studio

A crew of AI agents. One person has the final say.

What do we talk about tomorrow?The decision everyone faced.

What was the treatment?Now we can change one layer at a time.

Many of us are building AI games. Each design is a bet.

AI can write every word.What a learner must decide is still ours to design.

You Choose,
the LLM Narrates

Now it was fun.
That was the problem.

What can I sign off on?
The part that was always the design.

What do we talk about tomorrow?
The decision everyone faced.

What was the treatment?
Now we can change one layer at a time.

AI can write every word.
What a learner must decide is still ours to design.