Fostering Epistemic Insights into AI Ethics through a Constructionist Pedagogy: An Interdisciplinary Approach to AI Literacy

Ziyan Lin, Yun DaiAAAI 2025 · DOI

Introduction

As artificial intelligence (AI) becomes increasingly integrated into daily life, the need for AI literacy in K-12 education has grown. However, teaching the ethical dimensions of AI to young learners presents a significant challenge. Traditional methods often rely on abstract dilemmas, such as the "trolley problem," which can lead to a superficial understanding of the issues. This paper presents a constructionist pedagogical framework designed for primary school students (ages 10-12) that fosters "epistemic insights"—a fundamental understanding of how AI processes information—as a basis for ethical reasoning.

The Epistemological Gap in AI Literacy

A core obstacle in AI education is the prevalence of misconceptions among students. Many children view AI as either a "magic," human-like entity or a perfectly objective tool. Without understanding the underlying epistemology of AI—how it acquires and applies knowledge—students struggle to critically evaluate complex issues like algorithmic bias, transparency, and accountability. The proposed framework bridges the gap between technical "how-to" and philosophical "why-it-matters" by having students build and interrogate their own AI systems.

Methodology: Building to Learn

The researchers employed a constructionist approach, engaging 34 fifth-grade students in Hong Kong through three interdisciplinary activities combining computer science, mathematics, and ethics.

Activity 1: Rule-Based Classification

Students first created a "good/bad banana" sorter using Scratch. This activity required them to identify discrete features (e.g., color, size) and translate them into binary logic (1 for "yes," 0 for "no").

Students identify features like color and size to distinguish between different banana types.

By using explicit if-else logic, students explored the concept of transparency. They could see exactly why a banana was classified a certain way because they had programmed the rules themselves.

A Scratch script implementing a rule-based scoring system based on identified features.

Activity 2: Machine Learning Image Classifier

Moving beyond manual rules, students used Google Teachable Machine (GTM) to train an image classifier with a dataset of 60 banana images. This activity introduced the probabilistic nature of machine learning, focusing on confidence scores and the impact of biased training data.

Activity 3: Generative AI and Human Creativity

In the final activity, students used the Poe platform to generate poetry. By comparing the machine’s pattern-matching output to their own creative processes in Chinese language classes, students developed a deeper understanding of the distinctions between human and machine intelligence.

Key Insights: From Logic to Ethics

The study found that technical engagement served as a bridge to ethical reflection. The "aha moment" occurred when students realized that AI does not "think" but rather processes statistical patterns and binary code. This epistemic insight was crucial for debunking the "AI is magic" myth.

Furthermore, students experienced a "transparency trade-off." While the Scratch-based rules were clear and trustworthy, the ML models were often perceived as "opaque" or "black boxes." When a model misclassified a yellow banana, students used mathematical reasoning—such as probability and sample size—to justify their ethical concerns about fairness and accuracy.

Results and Ethical Implications

The constructionist approach yielded several key learning outcomes:

  • Fairness and Bias: Students identified how superficial features in training data (e.g., the color of a hand holding a banana) could lead to biased decisions.
  • Accountability: The responsibility for AI errors shifted from the machine to the developer. Students recognized their role as "data curators" who must prevent bias.
  • Human vs. Machine Intelligence: Students articulated that while machines require vast datasets to learn, humans can generalize from a few examples, highlighting a key epistemic distinction.
  • Real-World Application: Students successfully extrapolated from simple exercises to real-world harms, such as biased supermarket scanners causing food waste.

Conclusion

This research demonstrates that fostering epistemic insights through hands-on constructionist activities provides a robust foundation for AI ethics. While limited by a small sample size and the short duration of the Generative AI activity, the framework offers a promising model for interdisciplinary AI literacy. Future work should explore expanding this pedagogy to diverse educational contexts and older age groups.

Summary updated

BibTeX
@inproceedings{lin2025fostering,
  author = {Ziyan Lin and Yun Dai},
  title = {{Fostering Epistemic Insights into AI Ethics through a Constructionist Pedagogy: An Interdisciplinary Approach to AI Literacy}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35190},
}
Figure