BiasViz: A Project-Based, Narrative-Centered Learning Tool for Engaging Middle School Students in Critical Thinking about AI Biases

Paper Title

BiasViz: A Project-Based, Narrative-Centered Learning Tool for Engaging Middle School Students in Critical Thinking about AI Biases

Publication Info

• Topic area: AI ethics education for middle school students
• Keywords: AI bias, large language models, middle school education, project-based learning, narrative-centered learning, generative AI, critical thinking, AI literacy, educational tools, bias auditing

Background and Problem

• Problem / challenge: Few initiatives actively engage middle school students in identifying and quantifying biases in real-world generative AI systems, despite the importance of critical AI literacy.
• Significance: Understanding AI bias is crucial for fostering informed and equitable use of AI technologies, particularly as youth increasingly interact with AI-driven systems.
• Motivation and related work: Prior research has focused on auditing and mitigating AI bias, integrating ethics into K–12 AI education, and engaging students in co-designing future AI systems. However, little work has introduced middle school students to auditing biases in current generative AI systems like LLMs.

Solution

• Proposed approach: BiasViz, an interactive educational platform that integrates project-based and narrative-centered learning to help middle school students analyze biases in large language models (LLMs).
• Novelty:
  1. Combines project-based and narrative-centered learning to make AI bias analysis engaging and accessible for middle school students.
  2. Translates professional AI auditing practices, such as red-teaming, into simplified, student-friendly activities.
  3. Focuses on geographic bias in LLMs, leveraging students’ personal experiences to deepen understanding.
  4. Provides interactive tools for bias prediction, prompt engineering, and visualization of bias patterns.
• Procedure and key techniques:
  • Students brainstorm biases, test AI outputs, refine prompts, annotate biases, and visualize bias patterns using a seven-page workflow.
  • The platform uses OpenAI’s GPT-3 API for generating responses and analyzing next-word probabilities.
  • Learning objectives include defining biases, understanding their origins, practicing prompt engineering, and analyzing bias patterns.

Results

• Concrete findings:
  • 88% of students referenced at least one learning objective in their reflections.
  • 31% expressed interest in the app, 19% showed curiosity about AI bias, and 23% demonstrated critical thinking.
  • Post-activity assessments showed improvement in students’ articulation of AI bias concepts, with 52% providing correct examples of biased outputs and 32% designing effective bias-testing prompts.
• Advantage over baselines: Students with limited prior AI knowledge demonstrated foundational understanding and critical thinking about AI biases after using BiasViz. The platform successfully engaged students in hands-on bias auditing tasks.
• Experiments / evaluation:
  • Study conducted with 28 middle school students (grades 6–8) from a rural public school.
  • Two-hour sessions included pre-surveys, app interaction, reflection essays, group discussions, poster presentations, and post-surveys.
  • Data collected included interaction logs, open-ended responses, group artifacts, and assessments.
• Limitations and future work:
  • Short session duration limited depth of engagement; longer interventions are needed.
  • Ethics restrictions prevented audio-recording discussions, reducing qualitative detail.
  • Pre/post assessments were non-identical, limiting direct comparisons.
  • Future iterations will enhance scaffolding, reduce reliance on predefined examples, and integrate BiasViz into extended lesson plans.

Summary

BiasViz is an interactive tool designed to introduce middle school students to AI bias through project-based and narrative-centered learning. The platform engages students in hands-on tasks such as bias prediction, prompt engineering, and visualization, while encouraging them to draw on personal experiences. A study with 28 students demonstrated that BiasViz effectively fostered curiosity, interest, and critical thinking about AI biases, with many students articulating key concepts related to bias in LLMs. Future work will refine the platform and evaluation methods to deepen engagement and validate learning outcomes in broader educational contexts.