Uncovering Cultural Blind Spots: Divergences in Human vs. AI Interpretations of Culturally Responsive Mathematics Pedagogy

• In Event: Real insights from real classrooms: Motivation scholars’ experiences designing for and working with observational data

Abstract

Secondary mathematics teachers who serve Black, Brown and poor youth have historically struggled to turn key long-standing theories on culturally relevant-responsive pedagogies into actionable classroom practices. In response, Belonging-Centered Instruction (Matthews et al., 2021) is an observational framework born out of the voices of Black and Brown students through extensive longitudinal interviewing, underscoring seven dimensions of teacher practice that cultivate agentic and humanizing mathematics classrooms. In this study we focus on two of those dimensions, Decentering Teacher Authority (i.e., how teachers foreground students’ mathematical agency and expertise) and Communal Orientation (i.e., how teachers cultivate a classroom culture that centers interdependence above individual performance). By decentering teacher authority and embracing a communal orientation, mathematics teachers can disrupt conventional hierarchies that alienate historically marginalized students, instead cultivating ecosystems where student voice and cultural knowledge become indispensable to collective mathematical inquiry. When teachers cede intellectual space and foster interdependence, these two practices can ignite a self-sustaining cycle of belonging and empowerment (Gutiérrez, 2012; Matthews, 2021).
The cultural and education landscape, however, is being rapidly reshaped by advances in AI, prompting critical questions into whether machine-coded observations can compare to the lived expertise and intuition of human coders for capturing the richly contextual, culturally embedded practices that comprise transformative math instruction. Recent evidence shows that AI models reflect WEIRD (Western, Industrialized, Educated, Rich, and Democratic) bias, failing to grasp the characteristics of historically marginalized cultures due to limited cultural knowledge and/or inability to apply it appropriately (Atari et al., 2023; Mihalcea et al., 2025). While AI demonstrates comparable performances to humans in basic classroom observation tasks (i.e., generating summaries), it struggles in understanding cultural nuances and contextual subtleties that are essential for meaningful analysis (Amarasinghe et al., 2023; Koraishi & Karatepe, 2025). These cultural and contextual limitations in AI underscores the urgent need to study how such gaps materialize in culturally rich educational settings, particularly in mathematics classrooms serving Black and Brown students.
Thus, we present preliminary findings from our work that evaluates two essential research questions: 1) How do AI and human coders differ in evaluating Decentering and Communal practices in secondary mathematics instruction? 2) What cultural knowledge, experiences, and interpretive frameworks do human coders leverage that distinguish their coding from AI. This study utilizes a sequential mixed-methods design examining alignment between AI and human coders in identifying and interpreting belonging-centered instructional practices in over 300 secondary mathematics classroom videos. Phase one quantitatively compares human versus AI coding congruence across decentering and communal practices, analyzing agreement and predictive validity for teacher and student outcomes. The second, qualitative phase uses a case study approach, closely investigating observational events where humans and AI differ, particularly in culturally responsive practices, by engaging expert coders in focus groups, reflective analysis, and interpretive dialogue around their coding rationales and evaluation of AI’s reasoning. This two-pronged approach reveals not just where AI and human coding diverge, but what those differences uncover about the cultural knowledge and interpretive frameworks essential to cultivating belonging in mathematics classrooms.

Authors

• Jingyun Wu, Indiana University

• Jamaal Matthews, University of Michigan