Training Culturally Relevant AI Models for Science Instruction

• In Event: AI as a Co-Pilot in Science Education: Reimagining Instructional Practice

Abstract

Objectives
This study explored how a trained language model (LLM) supported culturally relevant science instruction through context-driven, conversational interactions. We investigated how the AI engaged students in personalized science discussions reflecting their local, linguistic, and cultural experiences. The goal was to examine how AI tools might enhance science teaching by modeling “teacher talk” that grounds content in students’ lived realities and scaffolds academic language development. We trained the model to evaluate, localize, and redirect like a science teacher.

Theoretical Framework
This work draws on Culturally Relevant Pedagogy (Ladson-Billings, 1995) and Pedagogical Content Knowledge (Shulman, 1987), emphasizing tailoring instruction to students’ cultural assets. The design incorporates sociolinguistics and language acquisition principles, using a “disaggregated instruction” model: scientific ideas are introduced in accessible, everyday language, then reframed with discipline-specific terminology. The AI emulates responsive, inquiry-based science instruction, with pedagogical strategies grounded in equity-oriented learning theories.

Methods
This mixed-methods study combined qualitative content analysis, student interviews, and survey data to assess the effectiveness of the CRP-trained LLM. One hundred transcripts of AI-student interactions in an augmented reality (AR) science environment were analyzed for how the AI identified cultural cues (local knowledge, language, practices) and adapted instructional explanations. Semi-structured interviews with 30 students probed their experiences with contextualized learning, language development, and perceived conceptual understanding. A post-survey measured students’ perceptions of learning via three constructs: culturally relevant learning, localized learning, and language scaffolding.

Data
Data included transcripts from 100 recorded AI-student instructional dialogues, interview data from a purposive sample of 30 middle school students, and survey responses from all participants. The data were collected during AR-based science activities, where students interacted with the AI model on core biology topics like mitosis, regeneration, and cellular function.

Results
Preliminary analysis suggests the AI model successfully identified cultural cues for teaching in all 100 interactions. The model used examples rooted in sports, cultural foods, and health care to build cultural bridges. It also leveraged family as a key to culturally meaningful connections. The AI’s ability to pivot explanations based on student context, such as linking mitosis to skin healing or cancer, fostered conceptual clarity and engagement. Transfer tasks (applying ideas to novel personal scenarios) revealed deeper understanding and agency. Interviews confirmed that using everyday language followed by academic terms boosted confidence and ownership of scientific discourse.

Scientific or Scholarly Significance of the Study
This study advances AI-mediated pedagogy by showing how large language models can embody culturally relevant teaching practices. By bridging local knowledge and scientific understanding, it offers a model for AI tools that listen, adapt, and teach in ways reflecting learners’ identities. The findings inform the design of equitable science education technologies and the future of teacher-AI collaboration in diverse classrooms.

Authors

• Bryan A. Brown, Stanford University

• Alan Cheng, Stanford University

• Lisa Teresa Archuleta, Stanford University

• Elizabeth Childs, Stanford University

• Roy D. Pea, Stanford University

• Laurence Tan, University of California - Los Angeles