AI as Archeological Evidence: Uncovering Dominant Teaching Discourse in Educational Technology

• In Event: AERA Poster Session Wednesday 11:45am
  In Poster Session: SIG Socio Political Issues in Mathematics and Science Education Poster Session

Abstract

This study uses an archaeological analysis of AI-generated feedback to examine how dominant teaching discourses are evident in the data used to train Large Language Models. When prompted to respond “as a teacher,” ChatGPT-4 reproduced evaluative feedback patterns that emphasize procedural correctness over conceptual reasoning. Analyzing 180 responses to student mathematical work, we draw on Apple’s (2018) hidden curriculum, Benjamin’s (2020) critique of coded bias, and Gee’s (2011) discourse analysis to uncover how AI reproduces patterns based on historically dominant pedagogical ideologies. Findings show that AI perpetuates one-directional approaches that prioritize correctness over reasoning. Findings suggest AI training data reflects historically dominant – yet pedagogically limited – conceptions of “good teaching,” raising questions about whose knowledge is privileged in AI training data.

Authors

• Cindy Yovanov, Drexel University

• Amanda Reinsburrow, Drexel University

• Jason Silverman, Drexel University