Bridging Pedagogy and Play: Introducing a Language Mapping Interface for Human-AI Co-Creation in Educational Game Design

This paper presents a web-based tool that utilizes a controlled natural language interface to enable non-expert educators and an LLM to collaboratively design educational games by explicitly mapping pedagogical intent to gameplay, thereby lowering design barriers while preserving human agency and ensuring alignment between learning goals and game mechanics.

The Gist

Imagine you are a teacher who wants to build a video game to help your students learn history or math. You have a brilliant idea for what you want them to learn, but you don't know how to code, and you aren't a game designer.

Usually, you'd have two bad options:

1. Do it yourself: Spend months learning to code, likely failing to make the game actually teach the lesson.
2. Ask an AI: Type "Make a game about the Civil War," and get back a finished game. But you have no idea how it works, if it's historically accurate, or if it's too hard for your students. You're just a passenger in a car driven by a robot you don't trust.

This paper introduces a new tool that solves this problem. Think of it as a specialized translator and co-pilot that sits between your teaching goals and the game design.

The Core Idea: A "Rosetta Stone" for Games

The researchers created a system where you and an AI don't just chat randomly. Instead, you both speak a very specific, structured language.

Imagine you are filling out a Mad Libs style sentence, but every word you pick has a specific job to do. The sentence looks like this:

"Players [Adverb] [Verb] [Noun] in a [Adjective] environment."

Here is what that means in plain English:

• Noun (The Topic): What are we learning? (e.g., "fractions" or "ancient Rome").
• Verb (The Action): What are the students actually doing? (e.g., "calculating" or "exploring").
• Adverb (The Rules): How well do they have to do it? (e.g., "quickly" or "carefully"). This sets the difficulty.
• Adjective (The Vibe): What does the world look like? (e.g., "futuristic" or "realistic").

How the Tool Works (The Three Steps)

The tool guides you through three phases, like building a house:

1. The Blueprint Phase (Requirement Extraction)
Instead of just saying "I want a game," the AI asks you specific questions to fill in that "Mad Libs" sentence.

• AI: "What specific skill should they learn?"
• You: "Adding fractions."
• AI: "How fast should they do it?"
• You: "Under 30 seconds."
• Result: The AI locks these choices into the sentence structure. This ensures you are the boss of the intent.

2. The Translation Phase (The Magic Bridge)
This is the coolest part. The AI takes your "Teacher Sentence" and instantly translates it into a "Game Sentence" using the exact same structure.

• Your Sentence: "Students quickly add fractions in a chaotic kitchen."
• AI's Game Translation: "Players rapidly combine ingredients in a messy kitchen."

The tool highlights the words in different colors so you can see exactly how "adding fractions" became "combining ingredients." If you don't like the "messy kitchen" part, you can tell the AI to change the "Adjective" slot, and it will instantly rewrite the game description to match. You aren't guessing; you are editing the blueprint.

3. The Construction Phase (Language Development)
Once you are happy with the sentence, you can click a "Zoom In" button. The AI expands that one sentence into a full paragraph describing the game, and then even writes pseudocode (a simple, human-readable version of computer code).

• Think of this as the AI handing you a detailed instruction manual and a rough sketch of the wiring, which you can then give to a professional game developer (or a coding tool) to build the real thing.

Why This Matters

Most AI tools are like ordering a pizza: you say what you want, and they give you a pizza. If you don't like the cheese, you have to throw the whole thing away and start over.

This tool is like cooking together in a kitchen.

• You hold the recipe card (the pedagogy).
• The AI is the sous-chef who knows how to chop and sauté (the game mechanics).
• But you are standing right there, tasting the sauce. If it's too salty, you say, "Add less salt," and the AI adjusts just the salt, not the whole dish.

The Bottom Line

This research shows that we don't have to let AI take over education. Instead, we can use AI as a translator that turns our teaching goals into game rules, while keeping us in the driver's seat. It makes sure the game actually teaches what we want it to teach, without needing to be a computer programmer to do it.

Technical Summary

Here is a detailed technical summary of the paper "Bridging Pedagogy and Play: Introducing a Language Mapping Interface for Human-AI Co-Creation in Educational Game Design."

1. Problem Statement

Educational games hold significant potential for fostering critical thinking and motivation, yet their adoption by instructors is hindered by a "design gap."

• Misalignment: Games often fail to align with specific course content, assessment practices, or classroom routines, leading to conceptual confusion and reduced pedagogical value.
• Skill Barriers: Instructors typically lack training in game design and learning experience design. While low-code tools exist, they do not resolve the deeper cognitive demands of mapping pedagogy to mechanics.
• AI Limitations: Current Large Language Model (LLM) assistants often generate "black box" solutions. Instructors receive seemingly complete games without insight into how design choices support learning objectives, leading to a loss of human agency. Instructors risk becoming mere content approvers rather than designers, unable to trace or critique the AI's reasoning regarding pedagogical alignment.

2. Methodology

The authors propose a Language Mapping Interface that operationalizes a controlled natural language framework as the central collaboration mechanism between human instructors and an LLM. The system is designed as a full-stack web application facilitating a three-phase co-creation workflow.

A. The Core Framework: Controlled Sentence Template

The system relies on a structured sentence template that serves as a shared representation for both pedagogical intent and game design. The template is:

"Players (Students) [Adverbs] [Verbs] [Nouns] in a [Adjectives] environment."

This template maps four grammatical elements to specific design dimensions:

1. Nouns (Content): The focal concept or content domain $\rightarrow$ Maps to Content models and in-game artifacts.
2. Verbs (Action): The observable target ability $\rightarrow$ Maps to Game mechanics and interaction patterns.
3. Adverbs (Performance): Performance requirements and success conditions $\rightarrow$ Maps to Rules, parameters, difficulty, and success conditions.
4. Adjectives (Context): Learning context, realism, and tone $\rightarrow$ Maps to Aesthetic profiles and narrative framing.

B. System Workflow (Three Phases)

The web application guides the user through three distinct phases:

1. Requirement Extraction:

   • The AI engages in a structured dialogue to elicit the four grammatical elements from the instructor.
   • The system asks specific questions to define the scope, materials, observable learner actions, performance targets, and contextual framing.
   • Instructors retain maximum authority to shape the design, using the game-language structure to refine their teaching intent.

2. Language Translation:

   • The AI generates a single "Pedagogy Sentence" based on the extracted requirements.
   • The AI then generates multiple candidate "Game-Language Sentences" that satisfy the same constraints.
   • Visual Mapping: The interface uses color-coding to link corresponding grammatical elements between the pedagogy sentence and the game sentence.
   • Rationale: For each candidate, the AI provides a component-wise rationale explaining how specific game elements correspond to pedagogical specifications, allowing instructors to compare, revise, or reject alternatives.

3. Language Development:

   • Instructors can iteratively edit the game-language sentence with AI assistance.
   • "Zoom In" Feature: Users can click to expand specific parts of the sentence.
     • First Zoom: Generates descriptive paragraphs with play examples to clarify game operations.
     • Second Zoom: Generates pseudocode for the game logic, serving as a reference for downstream development.
   • The final output is not a playable game but a structured design specification and pseudocode.

C. Technical Implementation

• Architecture: Full-stack web application.
• Frontend: Built with TypeScript and Node.js, featuring interactive, color-coded slot alignment and in-place editing.
• Backend: Implemented in Python for data orchestration, prompt construction, and logging.
• Database: MongoDB stores project data, including requirement responses, versioned sentences, and artifacts.
• AI Model: Utilizes the GPT-5 API (without fine-tuning). The system relies on structured prompting to constrain outputs to the controlled template and phase-specific objectives, ensuring consistency and reducing hallucinations.

3. Key Contributions

• A Controlled Language Framework: The paper introduces a novel sentence template that explicitly bridges the semantic gap between abstract pedagogical goals and concrete game mechanics, making the mapping process transparent and editable.
• Human-AI Co-Creation Interface: Unlike standard generative tools that output a final product, this interface treats the translation process as the primary artifact. It allows instructors to inspect, critique, and iteratively adjust the AI's proposals at a granular level (slot-by-slot).
• Traceability and Agency: By maintaining a visible, color-mapped link between teaching language and game language, the tool preserves human agency. Instructors can trace design decisions back to their original pedagogical intent, mitigating the "black box" risk of LLMs.
• Actionable Output: The system generates pseudocode and detailed design descriptions rather than just text, providing a tangible bridge to actual game development.

4. Results

Note: As this is an Extended Abstract (CHI EA '26) describing a system design and proposal, the paper does not report empirical user study results or quantitative metrics yet.

• System Status: The authors report the successful implementation of the full-stack web application and the functional workflow of the three phases.
• Design Validation: The paper argues theoretically that the framework addresses the identified problems by making pedagogical intent explicit and editable, thereby lowering design barriers for non-experts while preserving control.
• Future Evaluation: The authors outline a planned A/B evaluation comparing their language-mapping framework against standard LLM prompting, as well as studies on instructor trust calibration, mental model formation, and perceived ownership.

5. Significance

This work represents a significant shift in how AI is integrated into educational design:

• From Automation to Augmentation: It moves away from AI as a replacement for the designer toward AI as a collaborative partner that requires human oversight and intent specification.
• Pedagogical Transparency: It solves the critical issue of "alignment" in educational games by forcing a structured translation of learning objectives into game rules, ensuring that the resulting game mechanics are theoretically grounded in the intended pedagogy.
• Scalability for Non-Experts: By abstracting complex game design into a controlled natural language, the tool empowers instructors without game design backgrounds to create high-quality, aligned educational games, potentially democratizing the creation of interactive learning environments.