Stop! Yield! Do Not Enter: A Web-Based Self-Driving Vehicle Simulator for Building AI Literacy and Trust

Abstract

This innovative practice full paper presents an accessible, interactive approach to foundational AI literacy by visually demonstrating the consequences of imbalanced, biased, and poisoned training data on AI-based systems. As AI becomes increasingly integrated into critical domains such as transportation, public trust depends on understanding its limitations, necessitating basic AI literacy. To address this, we developed a web-based self-driving vehicle simulation that displays a twodimensional city with traffic signs at key intersections, powered by TensorFlow models trained using Google's Teachable Machine. The simulation provides real-time visual, situated, and gamified feedback on model performance, reinforcing the real-world implications of flawed datasets on AI behavior and decision-making. In a workshop involving 20 undergraduate and graduate students from the University of Massachusetts Lowell, participants used the simulation to examine how dataset flaws impact AI systems. Through hands-on interaction with intentionally flawed datasets, they identified issues, observed their effects, and iteratively improved model performance. Results showed that 60 % of participants achieved perfect accuracy, while the remaining 40 % misclassified only one or two images. Participants reported a deeper understanding of how dataset quality impacts the ethical and technical reliability of AI systems, along with a positive workshop experience and a strong willingness to recommend the activity to peers.

Ähnliche Arbeiten