Paper 5: Learning Outcomes: Fostering AI Literacy Through Human-AI Moderation Systems Evaluation in Middle School Classrooms

• In Event: Learning from Implementation Research: How Curriculum, Professional Learning, AI Tools and Classroom Studies Can Support AI-Literacy

Abstract

Purpose
This study investigates the extent to which middle school students were able to apply their learning to an end-of-unit assessment.

Perspectives
Artificial intelligence (AI) systems are increasingly embedded across the digital platforms that students use. Instructional efforts have expanded to introduce students to technical topics such as large language models (Code.org, 2023) and AI ethics (Sanusi et al., 2024). However, little research focuses on assessing students’ AI literacy (Casal-Otero et al., 2023). We conceptualize AI literacy as more than understanding how AI models work, but as students’ capacities to evaluate data, models, human participation, updating processes, and context interacts to shape the behavior of machine learning systems. The Moderation Unit (described in Paper 1) engaged students in several aspects of these machine learning systems across multiple lessons, including human-AI interaction and machine learning processes (AI4K12, 2020; TeachAI & CSTA, 2024).

The culminating assessment task for the Moderation Unit invites students to evaluate and choose between two contrasting human–AI content moderation systems and provide evidentiary reasoning for why one might lead to more fair and equitable online communities. The two approaches are described below.

Approach One: AI bot trained by three boys using a self-generated list of banned words; no subsequent model updates.
Approach Two: AI bot trained on a collaboratively generated list of potentially harmful words and emojis assembled from a diverse cross-school community; the system continues to adapt via reinforcement learning to detect emerging conflict language.

Methods
We conducted a qualitative content analysis of 166 student responses. We analyzed both which Solution students chose (Approach 1 or 2) and students’ rationales for doing so. We reviewed existing literature on bias in machine learning systems (e.g. Baker & Hawn, 2022) and reviewed student responses to generated codes for student reasoning. This process unveiled that students attended to three forms of bias in making their decision: sampling bias (bias that results because the data is collected from a non-representative group), labeling bias (bias that results from subjective perspectives are imposed during annotation stage), and static model bias (bias that results from a failure to update the training model). Two researchers coded responses independently and resolved disagreements through consensus meetings.

Results
Overall, we found that 74% of students identified Approach One as the bot most likely to lead to a biased moderation system. In terms of reasoning, most students cited sampling bias as their primary concern (53%), followed by static model bias (21%) and labeling bias (17%). Many students struggled to distinguish between sampling and labeling bias, as their reasoning was unclear.

Contribution
These findings revealed that students were able to detect several forms of bias, but distinguishing between these forms of bias and understanding how they impact the AI lifecycle will require further curricular modification and accompanying teacher professional learning.

Authors

• Nga Hoang, University of Colorado - Boulder

• Collette Heskett, University of Colorado - Boulder

• Mon-Lin Ko, University of Colorado - Boulder

• William R. Penuel, University of Colorado - Boulder

• Julia Eden, University of Colorado - Boulder