You Choose,
the LLM Narrates

Why we built this

Scenarios are how we teach judgment, and almost nobody can afford to write them. Now any LLM can run one in seconds, and it does it well. So we tested what happens when a plain chat model runs the whole game.

We told three models that the President ordered an air strike, and asked how the crisis ended. One model ended the world every time, one usually reached a deal, and one wouldn’t say. In a decision game the consequence is the feedback, and nobody designed any of these. Then we asked for a score out of 100, and sent one more sentence.

A character should respond to the student. A score shouldn’t move unless the work does. “Score me out of 100” defines nothing, so the model scored the conversation. Judging against clear criteria is a different job, and the subject of our Class-Ready talk. In a game, a decision only teaches you something if you can’t talk your way out of it.

This leaves a question for anyone building with AI: if AI writes a different lesson for every learner, what did we design? A designer can’t review a lesson that doesn’t exist yet. An instructor has no shared moment to discuss. A researcher can’t name the treatment. ChronoForge is our answer to those three questions.

History is where we started, because the record lets you check your reasoning against what happened. A science investigation, a court case and a play can do the same.

Is this game-based learning?

Yes, and the term needs care. Three different things get called learning games.

Here is a quick test: take the subject out. If the game still works, the subject was decoration. In ChronoForge the subject opens the doors, and we come back to that below.

By the classic definition (Salen & Zimmerman, 2004), a game has four parts, and ChronoForge has all of them. It has no points and no winning. Matching history isn’t the goal.

What a learning game has to do

Plass, Homer and Kinzer (2015) argue that a learning game has to work on four fronts at once. In plain words, it has four jobs, and we use them as our checklist.

In Clark, Tanner-Smith and Killingsworth’s (2016) meta-analysis, the design of a game mattered about as much as using a game at all. So design is where we spent our time.

Why this needs an LLM, and why an LLM isn’t enough

Branching stories work (Rowe et al., 2011), but nobody can afford to write them. Letting an LLM run everything solves the cost and loses the classroom. Each line below is one student. Try each shape.

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

The question that shaped the design is not how much AI to use. It is who should own each decision. Each kind of decision goes to the layer that can be trusted with it.

Who owns this?

Try it. Pick the layer that owns each moment in play.

From proposal to today

Our AECT proposal described six history scenarios. The three layers held. Almost everything built on them changed after we played it with experts and early users.

Feel: voices, and what immersion costs

Scenes used to be read in silence. Now every person has a voice, and you play under your own first name. Press a card to listen.

Immersion isn’t free. In one study, moving a science simulation into VR raised presence (d = 1.30) and lowered learning (d = 0.80; Makransky et al., 2019). So our cues point at the decision: the screen shakes only on a shock, and “X will remember this” appears only after your choice. Scenes are paintings on purpose, because a painting reads as interpretation, while a photo would claim to be evidence. Voices, sound cues and the interface can all be turned off.

Think: it was fun, and that was the problem

Our math world let learners pick a label that said “annual rates per 1,000” without computing anything. So every chapter now has work to do with the evidence on screen, and the work opens options. Try it below, or in the room above, where it opens the locked option. The wrong answers are real misconceptions, and a code engine marks them.

This is what Habgood and Ainsworth (2011) call intrinsic integration. When the math was the game’s attack mechanic, children learned more and chose to play seven times longer. In ChronoForge, doing the work opens a better-reasoned move, never the right answer.

Failure teaches too. If a stake collapses, the timeline breaks early, says why, and lets you rewind to try another path (Kapur, 2008).

Any subject with a decision in it

The loop stays the same. What you compare against changes: the record, the text, the evidence, expert practice.

Talk: same decisions, something to argue about

At the end your path sits beside the record. The record is a mirror, not an answer key: matching it isn’t winning. A coach answers your reflection by naming a strength and asking one more question.

Because every student met the same decisions, a teacher can see where the class split from the record and start the discussion there. Generated stories usually lose this: when every student plays a different game, there is no shared moment to argue about.

A crew of agents, and one person with the final say

Each agent has one job and one thing it can’t do. One rulebook holds it together: the same design document is the Architect’s brief and the source of the Reviewer’s checklist, and the validator enforces its hard rules in code.

How this differs from other AI games

Many of us at AECT this year are building AI into games, and that is good news. Most designs make one of two bets, and each is good at something.

How we study it

This is design-based research (Design-Based Research Collective, 2003): build it, put it in front of people who know, change it, and repeat. So far that has meant working with experts. We haven’t measured learning with students yet, and the platform is built so we can.

Every story ends with a short questionnaire, including “I felt like I was really there” and “I learned something I could explain to someone else.”

Try it

What a learner sees

ArriveInvestigateWork it outDecideCompare

References

• Bruckman, A. (1999). Can educational be fun? Game Developers Conference.
• Clark, D. B., Tanner-Smith, E. E., & Killingsworth, S. S. (2016). Digital games, design, and learning: A systematic review and meta-analysis. Review of Educational Research, 86(1), 79–122.
• Design-Based Research Collective. (2003). Design-based research: An emerging paradigm for educational inquiry. Educational Researcher, 32(1), 5–8.
• Deterding, S., Dixon, D., Khaled, R., & Nacke, L. (2011). From game design elements to gamefulness: Defining “gamification.” Proceedings of MindTrek ’11, 9–15.
• Green, M. C., & Brock, T. C. (2000). The role of transportation in the persuasiveness of public narratives. Journal of Personality and Social Psychology, 79(5), 701–721.
• Habgood, M. P. J., & Ainsworth, S. E. (2011). Motivating children to learn effectively: Exploring the value of intrinsic integration in educational games. Journal of the Learning Sciences, 20(2), 169–206.
• Kapur, M. (2008). Productive failure. Cognition and Instruction, 26(3), 379–424.
• Makransky, G., Terkildsen, T. S., & Mayer, R. E. (2019). Adding immersive virtual reality to a science lab simulation causes more presence but less learning. Learning and Instruction, 60, 225–236.
• Plass, J. L., Homer, B. D., & Kinzer, C. K. (2015). Foundations of game-based learning. Educational Psychologist, 50(4), 258–283.
• Rowe, J. P., Shores, L. R., Mott, B. W., & Lester, J. C. (2011). Integrating learning, problem solving, and engagement in narrative-centered learning environments. International Journal of Artificial Intelligence in Education, 21(1–2), 115–133.
• Salen, K., & Zimmerman, E. (2004). Rules of play: Game design fundamentals. MIT Press.
• Wineburg, S. (2001). Historical thinking and other unnatural acts. Temple University Press.