💻 computer science

AI Unplugged: Embodied Interactions for AI Literacy in Higher Education

This paper presents a novel pedagogical approach for higher education that integrates embodied, unplugged physical activities to foster holistic AI literacy by helping university students build intuitive understanding of complex concepts before transitioning to mathematical formalizations and code.

Original authors: Jennifer M. Reddig, Scott Moon, Kaitlyn Crutcher, Christopher J. MacLellan

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jennifer M. Reddig, Scott Moon, Kaitlyn Crutcher, Christopher J. MacLellan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach someone how to drive a car.

The Old Way (Traditional AI Classes):
You sit them in a classroom, hand them a thick textbook full of physics equations, and make them memorize the formula for friction and the mechanics of an internal combustion engine. You explain the theory of steering, but you never let them touch the wheel. By the time they finally get to drive, they are terrified because they understand the math of driving, but they have no feel for how the car actually moves.

The New Way (This Paper's Approach):
Instead of starting with the engine, you put them in a giant, empty parking lot. You give them a steering wheel and a map. You say, "Okay, turn left. Now, what happens? The car slides a bit, right? Now try to stop." They learn by doing, by feeling the car, by making mistakes and seeing what happens. Only after they've felt the car slide do you hand them the physics textbook to explain why it happened.

This paper, "AI Unplugged," is about bringing that "parking lot" experience to university-level Artificial Intelligence classes.

The Problem: AI is Too Magical

Right now, many students think AI is like magic. They see it in movies or on TikTok and think, "It just knows things!" But when they get to college, they are suddenly asked to write complex code and solve hard math problems. They get stuck because they don't understand the logic behind the magic. They are trying to build a house without ever having seen a brick.

The Solution: "Unplugged" Activities

The authors (teachers from Georgia Tech) decided to stop using computers for the first part of the class. They used "Unplugged" activities. This means:

No computers.
No code.
Just people, paper, cards, dice, and their bodies.

They treated the students like the AI agents themselves. Instead of writing a program to solve a maze, the students became the program.

The Four "Games" They Played

Here is how they turned boring math into fun, physical games:

1. The Search Game (Becoming Search)

The Concept: How does a computer find a path from point A to point B?
The Game: The class splits into groups of four. Each person holds a secret piece of information on a card (like a map fragment). One person is the "Algorithm," one is the "Goal," one is the "Map," and one is the "Memory." They have to pass cards back and forth to find a hidden treasure.
The Lesson: Students realized that an AI doesn't see the whole map at once (like a human looking at a map). It only sees one step at a time. By physically passing cards, they felt the frustration and logic of searching step-by-step.

2. The Card Game (Red and Black Jack)

The Concept: Making decisions when the future is uncertain (Markov Decision Processes).
The Game: A version of Blackjack. But instead of trying to get 21, you are trying to guess if you should "Hit" (take a risk) or "Stand" (play it safe) based on the cards you've seen.
The Lesson: Students learned that AI has to weigh risks. If they draw too many "bad" cards, they lose points. They physically felt the difference between a "safe" strategy and a "risky" one, which helped them understand how AI calculates the best move when it doesn't know what will happen next.

3. The Human Maze (Q-Learning)

The Concept: Learning by trial and error (Reinforcement Learning).
The Game: Students walked through a grid drawn on the floor. Each square had a piece of paper with a "reward" written on it. They rolled dice to decide whether to explore a new square or stick to the path they knew was good.
The Lesson: They acted as the AI "learning" the maze. At first, they were lost (exploring). As they played more rounds, they started remembering which paths gave them the most points (exploiting). They physically experienced the "Aha!" moment when the AI finally figures out the best route.

4. The Spy Game (Hidden Markov Models)

The Concept: Guessing where something is when you can't see it clearly (like a robot trying to find a person using noisy sensors).
The Game: One student is a "Spy" hiding in a city. Another is a "Hunter." The Spy moves secretly, but the Hunter gets vague clues (like "The spy is near the park"). The Hunter has to guess where the Spy is based on those fuzzy clues.
The Lesson: Students learned that AI often has to make guesses based on imperfect information. They realized that even with bad data, you can narrow down the possibilities if you use the right logic.

Why This Matters

The teachers found that when students played these games:

They stayed awake: Attendance went up because it was fun.
They understood the "Why": When they finally went back to the computers to write the code, the math didn't feel like magic anymore. It felt like the rules of the game they had just played.
They built confidence: Students who were scared of math realized, "Oh, I already know how to do this; I just did it with cards!"

The Takeaway

This paper argues that to teach AI well, we shouldn't just dump code on students. We need to let them play with the ideas first.

Think of it like learning to swim. You don't start by reading a book about fluid dynamics. You start by getting in the shallow end, splashing around, and feeling the water. Once you aren't afraid of the water, then you learn the science of swimming.

By "unplugging" the classroom, these teachers helped students swim in the ocean of Artificial Intelligence without drowning in the math.

1. Problem Statement

The paper addresses a critical gap in higher education AI curricula: the disconnect between abstract algorithmic instruction and holistic AI literacy.

The Issue: Current university-level AI courses often prioritize code-heavy assignments, mathematical formalization, and algorithmic techniques while neglecting conceptual understanding, ethical reasoning, and the "first-person" experience of AI decision-making.
Student Misconceptions: Students often enter courses with "magical" or pop-culture-influenced views of AI. They struggle to reconcile media portrayals with the routine, deterministic, or probabilistic algorithms they are asked to code.
Pedagogical Limitations: Traditional lectures and project-based learning often lead to students grappling with implementation details (syntax, libraries) without a deep intuitive grasp of the underlying logic. Furthermore, K-12 frameworks (e.g., UNESCO) emphasize holistic literacy (ethics, data, reasoning), but higher education frequently fails to integrate these dimensions effectively.
The Barrier: High barriers to entry regarding mathematics and programming anxiety prevent many students from engaging deeply with AI concepts.

2. Methodology

The authors propose and implement a pedagogical framework based on Embodied Cognition and CS Unplugged principles, adapted for university-level instruction.

Core Philosophy: The curriculum utilizes "unplugged" activities (physical, collaborative, non-digital) to simulate AI agent behaviors. This allows students to experience the constraints, information distribution, and decision-making processes of AI agents firsthand before transitioning to mathematical formalization and coding.
Course Structure: The study was conducted during a 12-week Summer 2025 semester of an "Introduction to Artificial Intelligence" course (40 students).
- Cycle: Each topic followed a Unplugged $\rightarrow$ Plugged $\rightarrow$ Reflection cycle.
- Process:
  1. Unplugged: Students engage in physical games modeling specific AI concepts.
  2. Formalization: Instructors guide students to map the physical game mechanics to mathematical models (e.g., transition matrices, Bellman equations).
  3. Plugged: Students implement the algorithm in Python (Jupyter Notebooks).
  4. Reflection: Students analyze the ethical and practical implications of the algorithms.
Key Activities: Four specific embodied activities were designed to cover the course's four main pillars:
1. Search (Becoming Search): Students act as distributed roles (Successor Dictionary, Goal Test, Frontier, Algorithm) using index cards to simulate Breadth-First Search (BFS), Depth-First Search (DFS), Uniform-Cost Search, and A*. This highlights the difference between a "bird's-eye view" (human) and a "local state view" (agent).
2. Markov Decision Processes (Red and Black Jack): A card game simulating stochastic environments. Students play "Hit" or "Stand" with specific probabilities to learn about states, actions, transition probabilities, and value iteration.
3. Q-Learning (Q-Maze): Students physically traverse a grid on the floor, acting as a single pass of a Q-learning agent. They use dice to determine exploration vs. exploitation and update Q-values on whiteboards, visualizing the convergence of the Q-table.
4. Hidden Markov Models (Two Spies): A game involving a "Deep Cover Spy" (hidden state) and a "Hunter Spy" (observer). Students simulate transition and sensor models using dice rolls to update their "belief" state about the spy's location, introducing probabilistic inference.

3. Key Contributions

Novel Pedagogical Model: The paper presents a validated framework for integrating embodied, unplugged activities into university-level AI education, bridging the gap between K-12 conceptual learning and higher education technical rigor.
Curriculum Design: It provides four specific, reproducible activity designs that map physical interactions to complex AI concepts (Search, MDPs, Q-Learning, HMMs).
Bridging Mechanism: The authors detail a specific strategy for transitioning from intuitive, physical reasoning to mathematical formalization and code implementation, demonstrating how the physical constraints of the game motivate the necessity of the mathematical model.
Scalability Insights: The study offers a critical analysis of the feasibility of active learning in higher education, addressing challenges related to class size, time constraints, and student resistance to non-traditional instruction.

4. Results

High Engagement: Despite the short summer semester, class attendance remained high (75–80%) compared to previous lecture-heavy semesters (<30%). Students were observed actively discussing strategies and wins during activities.
Conceptual Transfer: Students successfully translated their physical experiences into algorithmic thinking.
- They intuitively understood the explore-exploit trade-off in Q-Maze without explicit prompting.
- They grasped the distributed nature of information in search algorithms, correcting the misconception that agents have global knowledge.
- They could manually perform value iteration and belief updates because the physical game mechanics made the abstract math concrete.
Confidence Building: The "low barrier to entry" approach (using dice/cards) reduced math anxiety. Students with weaker math backgrounds were able to engage with the logic of the algorithms before tackling the code.
Implementation Success: Students were able to write Python code that mimicked the roles and logic they had physically enacted, leading to more robust and conceptually grounded implementations.

5. Significance

Holistic AI Literacy: The approach moves beyond "how to code AI" to "how AI thinks," fostering a deeper understanding of limitations, uncertainty, and decision-making processes.
Democratization of AI Education: By lowering the initial barrier of complex mathematics and syntax, the method makes advanced AI concepts accessible to a broader range of students, including those with math anxiety.
Ethical and Critical Awareness: The embodied nature of the activities forces students to confront the "blind spots" of agents (e.g., lack of global view, reliance on noisy sensors), naturally leading to discussions on bias, uncertainty, and ethical deployment.
Future Directions: The paper suggests that while effective, these activities require careful scaffolding and time management. Future iterations should focus on multi-day structures to allow for deeper analysis of parameter changes and scaling these methods to larger university classes (300+ students) through better self-management strategies for student groups.

In conclusion, the paper argues that embodied, unplugged interactions are not merely introductory tools for K-12 but are essential, high-impact pedagogical strategies for university-level AI education, capable of fostering both technical proficiency and critical conceptual understanding.