Breakable Machine: A K-12 Classroom Game for Transformative AI Literacy Through Spoofing and eXplainable AI (XAI)

University-school collaboration (resource design for K-12 AI education)

Finland (designed for broader K-12 classroom use)

Researchers and teachers co-developing a classroom AI spoofing/XAI game

• In-school (K-12)

Classroom game with collaborative leaderboard and guided reflection

Classroom session integrated into broader AI project units

Whole-class parallel play with individual or small-group participation

Teacher projector app plus student camera-enabled devices via browser

• Learners need prior exposure to basic image-classification concepts.
• Activity effectiveness depends on teacher-facilitated reflection beyond gameplay.
• Hardware/network access and classroom orchestration can affect implementation quality.

Grades 4-9 (approximately ages 10-15)

Beginner-to-intermediate familiarity with classification concepts

No coding required for game participation

• Understand that high AI confidence does not guarantee correctness.
• Experience adversarial spoofing to reveal classifier fragility and failure modes.
• Develop critical reasoning about AI systems as human-made sociotechnical artifacts.

• Interpret feature saliency through XAI heatmaps.
• Discuss fairness, accountability, and real-world consequences of misclassification.
• Build learner agency to question and reshape AI-mediated practices.

• Not a model-building-only coding lesson.
• Not a benchmark study of classification accuracy improvements.
• Not a claim that spoofing itself is the final learning endpoint.

Browser-based spoofing game with XAI visual explanations and class leaderboard

Visual interaction and teacher-facilitated classroom discussion

• Evaluator

• Teacher launches challenge labels and class controls through projector interface.
• Students manipulate appearance/background to induce high-confidence misclassification.
• Learners inspect CAM heatmaps to infer which visual features drive predictions.
• Class compares top spoofing strategies and reasoning via shared leaderboard.

• Local-session data handling with no persistent identifiable storage outside classroom.
• Privacy-conscious design aligned with GDPR-oriented principles in deployment context.
• Teacher-mediated reflection to avoid normalizing unsafe or unethical AI misuse.

• Students join by QR code and receive target classification challenge.
• Students iterate spoofing attempts while observing confidence feedback.
• Students switch to XAI heatmap and training-data views for evidence gathering.
• Classroom discussion links observed failures to robustness, bias, and accountability.

• Humans design challenge conditions, interpret outcomes, and guide ethical inquiry.
• AI model produces probabilistic labels and saliency outputs.
• Learners critically evaluate model behavior and propose sociotechnical implications.

• Adversarial play reframes model failure as a pedagogical investigation target.
• Peer-visible leaderboard supports collaborative hypothesis testing and sensemaking.
• XAI visualizations make hidden model attention patterns discussable in class.

• Students may over-focus on game performance unless reflection is intentionally scaffolded.
• Conceptual transfer from spoofing tactics to ethical/systemic reasoning needs facilitation.
• Classroom logistics (device readiness, pacing, projection visibility) can constrain flow.

• Included dual interfaces (teacher control + student interaction) for classroom manageability.
• Added heatmap and training-data views to move from play to explanation.
• Used shared leaderboard to externalize successful strategies and foster collective inquiry.
• Positioned activity as part of broader pedagogical models (e.g., CEDE) instead of standalone use.

• Game format promotes high participation through playful challenge and peer comparison.
• Embodied experimentation supports active inquiry into model behavior.

• Learners can observe that semantically irrelevant visual cues can drive predictions.
• Students encounter and reason about confidence-vs-correctness mismatches.
• Activity opens pathways to discuss bias, accountability, and societal impact of AI errors.

This resource is best used as a bridge from technical curiosity to critical and transformative AI literacy, where failure analysis becomes core content rather than a side effect.

• Design emphasizes classroom-local data handling and automatic session-end deletion.
• Learners are prompted to examine who is harmed when AI misclassifies people in real contexts.
• Critical discussion targets representational harm, dataset assumptions, and governance implications.

• Activity documentation
• Practitioner observation

• Provides a concrete model for teaching AI robustness and adversarial behavior in K-12.
• Expands AI literacy pedagogy from model-building toward critical system interrogation.
• Offers a practical XAI-mediated entry point for discussing fairness and accountability.

• Transformative AI literacy
• K-12 XAI pedagogy
• Adversarial thinking in education
• Data agency and sociotechnical critique

• Completed

Grades 4-9 (10-15 years)

Classroom game-based AI inquiry

Image classification, saliency explanation, and spoofing analysis

Adversarial play with collaborative reflection and critical discussion

Medium

Low to Medium (in-session camera inputs and classroom-generated traces)