AI Unplugged: Embodied Interactions for AI Literacy in Higher Education

AI summary

Overview

Research area: AI education and computing education (cs.CY) — specifically AI literacy, active learning, and "unplugged" (technology-free) pedagogy in university settings.

Technical level: Intermediate. The paper is readable by anyone, but the content assumes familiarity with the AI topics it teaches (search algorithms, MDPs, Q-learning, HMMs), which are university-level.

Scope: An exploratory case study describing four embodied, unplugged classroom activities integrated into a 12-week university "Introduction to Artificial Intelligence" course, along with reflections on implementation challenges and proposed refinements.

What This Paper Is About

Most university introductory AI courses are lecture-heavy and code-centric, and students often arrive with pop-culture-driven notions of AI as "magical" or superhuman (Hingle and Johri 2023), making it hard for them to reconcile media portrayals with the routine algorithms they code in class. Building on the K-12 "CS Unplugged" tradition and the theory of embodied cognition, the authors designed physical, collaborative games that let students act out AI decision-making as agents themselves. The goal is to see whether unplugged activities can bridge conceptual reasoning and technical skill-building in higher education, not just in K-12.

Key Contributions

1. Four unplugged AI activities for a university course: Becoming Search (search algorithms), Red and Black Jack (Markov Decision Processes), Q-Maze (Q-learning), and Two Spies (Hidden Markov Models).
2. An explicit unplugged-to-plugged bridge: a described structure in which each activity is followed by student articulation of reasoning, a mathematical formalization, and then a Python/Jupyter Notebook implementation of an AI agent.
3. A viability probe for unplugged AI instruction in higher education, via a case study of a joint unplugged/plugged Introduction to AI course taught in the Summer 2025 semester.
4. A catalog of opportunities and challenges for university-level unplugged instruction, plus concrete proposals for future iterations of the course.

Main Findings

• Attendance and engagement rose sharply. Although attendance was not part of the course grade, sessions "routinely attracted 75-80% of students through the end of the semester," compared to "less than 25-30%" at the end of prior, more lecture-based semesters. Students discussed wins, decisions, and strategies loudly with each other, and every student contributed insight either verbally in the class-wide debrief or through a written personal reflection.
• Students connected intuition to formal algorithms. Groups in Becoming Search shared information to find a path and then coded helper functions mimicking each role; students in Q-Maze grasped the explore-exploit trade-off (hoping for the "explore" dice roll when the Q-table was full of zeros, preferring the maximum-valued action once entries filled in) and used that experience to tune parameters in their Q-learning coding project; students in Red and Black Jack expressed joy at seeing their predicted optimal policy emerge via value iteration; and in Two Spies students spontaneously used extra dice as tokens to represent belief states.
• Familiar probabilities made abstract parameters concrete. Because students could draw on dice and card probabilities, numbers that would appear arbitrary in a bare example like GridWorld instead appeared as obvious properties of the problem.
• Prior CS and math coursework was an advantage. Data Structures and Algorithms was a prerequisite, so most students already knew stacks, queues, multi-dimensional arrays, and algorithmic thinking, letting the instructional team spend less time on foundational math and programming and accelerate to more complex technical details.
• Content coverage and time were the main cost. The summer course runs 12 weeks instead of 16, and running the activities required extra time for setup, exploration, discussion, coding, and reflection. The authors argue that judging a course only by how much content is included may not be the best representation of student takeaways.
• Scale is an unresolved issue. The course had only 40 students, compared to Fall and Spring sections that can have up to 300 students in a single section; orchestrating the activities required significant involvement from the instructor and TAs.
• No quantitative learning-outcome measures are reported. The paper reports engagement observations and qualitative reflections; it does not report pre/post test scores, grades, or effect sizes. The authors state that they "probe the viability" of the approach and describe results as an exploratory case study.

Methodology in Plain English

The authors kept the course's existing structure — four main topics (Search; Reasoning with Uncertainty, covering Markov Decision Processes and Q-Learning; Probabilistic Reasoning through Time, covering Dynamic Bayes Nets, HMMs, and Filtering; and Machine Learning, covering Neural Networks, Decision Trees, Classification, and Regression) — and its four in-depth coding projects in Jupyter Notebooks. Students developed algorithms to navigate a house and find treasure, approximated an optimal policy through repeated trials, localized a hidden object, and built a neural net from the ground up. After each project, students reflected on the applied algorithms' uses, benefits, and potential issues through Socratic Mind (Hung et al. 2024).

For the Summer 2025 offering, the authors repurposed in-class lecture time for unplugged experiences, running one activity per class session and sometimes repeating an activity over several days to add complexity. Each activity used physical objects, paper, and pencil, with no computers, and often used familiar mechanics like dice and cards. Hidden information was used deliberately so students had to decide step by step rather than guess a solution from a bird's-eye view. After each activity, students articulated their reasoning and strategies, the class worked through a mathematical formalization motivated by the activity, and then students built a Python agent using that formalism.

The four activities were:

• Becoming Search: Groups of four, each student taking one role — Successor Dictionary, Goal Test, Frontier, and Algorithm — with information distributed across index cards. The setup initially models Breadth-First Search (a queue in the Frontier); swapping to a stack models Depth-First Search, adding action costs and a priority queue models Uniform-Cost Search, and adding a heuristic to the Goal Test with estimate-ordered states models Greedy or A* Search. Students swap roles and repeat on a new scenario.
• Red and Black Jack: A single-player card game played in pairs using a standard 52-card deck, with a "Hit" deck (two red cards, two black cards, one face card) and a "Stand" deck (one red card, one black card). Scoring: bust earns −5 points, Single earns 1 point, Double 5 points, Triple 15 points, and Jackpot 30 points. The formalization has 9 non-terminal states and 5 terminal states, which is small enough that students can perform value iteration by hand; the algorithm converges within a few iterations, and rewards were chosen so the end values would be "nice" numbers.
• Q-Maze: A grid taped out on the floor (3x3, 4x4, 3x4, 5x5, etc. as space allows), where each cell is a state with a whiteboard showing its reward and the q-values for available actions. One episode: flip the starting position's board, roll a die to choose explore vs. exploit, flip the chosen action's board, use the Bellman Equation to update the q-value, flip back the previous state, and repeat until a terminal state or until actions run out. The number of allowable actions is the grid's length plus its width. Different grids let students navigate when q-values are all zero and later when they have mostly converged; the class then debriefs on the converged Q-table.
• Two Spies: Inspired by the mobile game, students in pairs play a Deep Cover Spy (which simulates the HMM) and a Hunter Spy (actions: Move, Stay, or Capture). The Deep Cover Spy rolls dice to determine state transitions and again to determine observations; making observations refer to regions rather than individual cities led to more interesting inferences. Players get six rounds to capture or evade, and students swap roles after three games.

Materials are available at https://github.com/jmreddig/Intro-to-AI-Teaching-Resources. The work was supported by the National Science Foundation under Grant No. 2247790 and Grant No. 2112532.

Why This Matters

Impact on research. The paper argues that although CS Unplugged is well studied in K-12 and AI Unplugged activities have been explored there (for example by Lindner et al. 2019), unplugged activities are rare in higher education. Lindner et al. (2019) noted that connecting student thought processes to the ways a machine learns makes AI easier for K-12 students to comprehend; this paper suggests that finding holds in higher education as well, and offers a model for a "unplugged to plugged to reflection" learning cycle aligned with the UNESCO framework.

Real-world applications:

• University AI and machine learning courses that want to reduce math anxiety and low self-efficacy barriers (Allen et al. 2021) while still teaching rigorous content.
• Introductory data science, robotics, or reinforcement learning courses that use GridWorld-style problems, which the authors note can demonstrate search, action stochasticity, sensor stochasticity, and robot navigation challenges in one flexible scenario.
• Outreach, bridge, and pre-college programs that need low-equipment, self-contained activities runnable in a classroom with paper, dice, and cards.
• Corporate or professional training where learners need intuition about AI decision-making and its limitations before touching code.

Industry relevance. As AI becomes embedded in social media, music recommendation, and generative text and images, the authors argue people need to understand how these systems work, what they can and cannot do, and when a situation actually calls for an AI solution — particularly the people who design, build, and deploy them. The paper's framing that students can learn to not just implement AI systems but understand and evaluate them in real-world contexts speaks directly to industry demand for practitioners with judgment about AI, not only coding fluency.

Future Directions

1. Make reflection universal and explicit. Not all students vocalized their reasoning, so the authors propose asking all students to outline their procedure in writing or verbally. In Two Spies specifically, they would walk through the matrix multiplications by hand step-by-step — something they did not fully unpack — and have students propose an algorithm for combining transition and observation probabilities to update belief.
2. Apply multiple AI methods to each scenario. For example, hiding the deck contents and rewards in Red and Black Jack to teach Q-learning through approximated transition probabilities and rewards, or applying Particle Filters to Two Spies using small tokens like M&M's or Skittles to approximate particles. GridWorld is highlighted as flexible enough to span search, action stochasticity, sensor stochasticity, and combined robot navigation challenges.
3. Spread activities across multiple days. The Search activity's variants ran over two days, which the authors say limited attention to each change (frontier structure, action costs, heuristics). A multi-day structure would let students internalize mechanics and take on added complications, such as exploring the effect of discount factors, reward structures, transition probabilities, exploration rate, and learning rate.
4. Scale the approach to larger classes. The authors identify scaling as an open problem, given the 40-student summer roster versus sections that can reach 300 students, and hope their self-contained activities can be self-managed by student groups.

Target Audience

This paper is most useful for university instructors and curriculum designers teaching AI, machine learning, or computational courses who want active-learning alternatives to lecture and code-heavy assignments; computing education researchers interested in extending CS Unplugged and embodied cognition into higher education; and K-12 AI curriculum developers curious about how their frameworks continue at the university level. Teaching assistants and course staff who help orchestrate activities will also benefit from the practical setup details and the candid discussion of challenges around scale and content coverage. Readers looking for quantitative evidence of improved learning outcomes will not find it here — the paper is an exploratory case study whose evidence is engagement data and qualitative observation, and its authors present it as a starting point rather than a validated intervention.

Authors’ abstract

As artificial intelligence (AI) becomes increasingly integrated into daily life, higher education must move beyond code-centric instruction to foster holistic AI literacy. We present a novel pedagogical approach that integrates embodied, unplugged activities into a university-level Introduction to AI course. Inspired by the effectiveness of CS Unplugged in K-12 education, our physical, collaborative activities gave students a first-person perspective on AI decision-making. Through interactive games modeling Search Algorithms, Markov Decision Processes, Q-learning, and Hidden Markov Models, students built an intuition for complex AI concepts and more easily transitioned to mathematical formalizations and code implementations. We present four unplugged AI activities, describe how to bridge from unplugged activities to plugged coding tasks, reflect on implementation challenges, and propose refinements. We suggest that unplugged activities can effectively bridge conceptual reasoning and technical skill-building in university-level AI education.