UC Berkeley

This dissertation investigates how AI adaptive dialogs can be designed, iteratively refined, and customized in partnership with teachers to support middle school students' integration of complex science and social justice ideas in culturally and linguistically diverse classrooms. A central argument is that teacher expertise within a Research-Practice Partnership (RPP) operates at two distinct levels: through collective co-design, teachers shape AI dialogs that work broadly across diverse classrooms; through local customization, individual teachers amplify what those dialogs can accomplish for specific students in specific community contexts. These two levels form an iterative refinement in which local observations feed back into collective improvement, generating design knowledge that neither level could produce alone..The Knowledge Integration (KI) framework and Justice-Centered Science Pedagogy (JCSP) provided a merged theoretical foundation for dialog design, instruction, and analysis. KI guided rubric development, adaptive guidance design, and student learning assessment through a five-level scoring rubric that rewards linked reasoning. JCSP extended KI's commitment to honoring diverse student ideas to include community observations and structural critique of systemic inequity. Together, the frameworks shaped how the RPP built the dialogs, supported teacher customization, and measured what students learned. The dissertation draws on five years of design-based research across six socioeconomically and racially diverse schools in the western United States, involving more than 5,000 student responses analyzed through cumulative link mixed models, generalized estimating equations, and McNemar tests, alongside qualitative analysis of teacher interviews and classroom observations. Chapter 2 establishes that iterative teacher-informed rubric expansion improved both NLP model accuracy and student KI gains across four design cycles of the photosynthesis dialog. Chapter 3 shows that two teachers' distinct pedagogical goals, making ecosystem connections visible and making science accessible for emergent bilingual learners, shaped dialog refinement and classroom instruction in complementary ways, both producing significant learning gains. Chapter 4 extends this work to food justice, showing that RPP co-design can build effective AI dialogs for social justice science topics, and that teachers' pedagogical orientations toward food justice shaped which ideas became available in students' reasoning.Together the three chapters show that AI dialogs become most effective when teacher expertise is embedded in the tools themselves through sustained co-design, and when teachers design instruction that treats the dialog as a starting point rather than a finished product. This dissertation offers the field actionable design guidelines and empirical evidence for building AI dialogs that honor the full range of student thinking, including community-based, culturally grounded, and justice-oriented ideas, and that are worthy of the students and teachers they are designed to serve.