Abstract
Data bias in artificial intelligence (AI) and machine learning (ML) presents significant challenges across sectors like healthcare and criminal justice. As these technologies become more pervasive, educating youth on the nuances and implications of data bias is essential. This paper introduces SourceML, an approach that teaches data bias by engaging high school students in building ML models using historical primary sources. Utilizing the web-based tool StoryQ, students analyze text data within a history classroom context. Our results indicate that the SourceML approach enhances students' understanding of bias origins, societal impacts, and mitigation strategies, emphasizing the role of historical thinking in data literacy.
Introduction: The "Black Box" Problem
While AI and ML increasingly govern critical aspects of modern life, education regarding algorithmic bias often remains siloed within technical STEM curricula. This creates a gap where students may understand the mechanics of code but lack the interdisciplinary skills to interrogate why data is biased. The authors argue that data bias is rarely a mere technical glitch; it is frequently rooted in historical inequalities and human subjectivity. To address this, the SourceML approach integrates AI literacy into history classrooms, using "historical thinking"—the practice of questioning a source's origin, context, and perspective—as a framework for auditing datasets.
The SourceML Framework
The researchers implemented a five-hour curriculum module using StoryQ, a no-code machine learning and text-mining tool designed for K-12 education. The pedagogical workflow is structured around four distinct phases that bridge historical inquiry with data science.
Pedagogical Phases:
- Problem Statement: Students examine the history of "redlining"—discriminatory 1930s lending practices by the Home Owners Loan Corporation (HOLC).
- Exploration of Model: Using StoryQ, students build a Unigram model to classify historical neighborhood descriptions into "A" (desirable) or "D" (hazardous) categories.
- Feature Analysis: Students investigate which specific words (features) the model prioritizes to make its classifications.
- Discovery & Debrief: Students compare model predictions against modern "counternarratives" they write themselves, highlighting what the original historical data omitted.
Datasets and Implementation
The curriculum utilized two primary data sources:
- HOLC Redlining Dataset: Approximately 400 primary source descriptions from the late 1930s that served as the basis for systemic housing discrimination.
- Student-Generated Counternarratives: Modern descriptions of communities (e.g., Bronzeville) written by students to serve as testing data, providing a contemporary lens to contrast with historical prejudices.
Key Insights: Historical Thinking as Data Literacy
The study revealed that the skills used to analyze a 19th-century diary are directly transferable to auditing a modern dataset. Key takeaways included:
- Coded Language as Features: Students discovered the model used seemingly neutral words like "trees" or "wooded" as strong predictors for wealthy, white neighborhoods. This helped them realize how language can mask racial and socioeconomic bias.
- Bias as a Human Product: Students shifted from viewing AI as an objective authority to seeing it as a reflection of its creators. As one student noted, "It depends on what they feed it," acknowledging that human judgment is baked into training data.
- Multidisciplinary Solutions: Participants began to advocate for solutions beyond technical fixes, recognizing that mitigating bias requires ethical frameworks and diverse perspectives during data collection.
Results and Impact
The SourceML intervention yielded significant quantitative and qualitative improvements in student literacy:
- Quantitative: Pre- and post-assessments showed a statistically significant increase in students’ understanding of data bias ($U = 462, p < 0.05$).
- Qualitative: 71% of interviewed students (15 out of 21) demonstrated a sophisticated grasp of bias origins, such as historical inequality and data labeling.
- Societal Awareness: Students successfully generalized their learning, connecting historical redlining to modern issues like biased search engine results and recommendation algorithms.
Conclusion and Future Directions
The SourceML approach demonstrates that historical primary sources provide a rich, contextualized foundation for teaching AI ethics. While this initial case study was small-scale (42 students), it provides a roadmap for scaling interdisciplinary AI education. Future work will focus on testing the approach in diverse geographic settings, expanding to other datasets like modern consumer reviews, and introducing mathematical metrics for quantifying bias, such as disparate impact and demographic parity.
Summary updated