Interactive Machine Learning and Conversational Agents for Supporting Scientific Inquiry in Secondary Education

• In Event: Innovating K–12 Teaching and Learning with Generative AI: Insights From IES-funded National R&D Centers

Abstract

Objectives
Preparing secondary students for authentic scientific and engineering practices, including inquiry, data literacy, and engagement with AI, is a pressing educational goal (Wang, 2025). InquiryAgent was developed to empower students and teachers to participate in data-driven inquiry, modeling, and reflection using accessible machine learning (ML) and generative AI (GenAI) technologies. By engaging students in InquiryAgent, they will further their competence in scientific practices, promote agency and critical thinking, and provide adaptive scaffolding for metacognition and access to AI education. This study aims to investigate how students engage in AI-based scientific inquiry to learn science and how the experience improves their interest in science.

Perspectives
InquiryAgent is grounded in the urgent need to align science education with how AI is transforming scientific inquiry (Author, 2024). Anchored in constructionism and three-dimensional learning (National Research Council, 2012), our design positions students as creators of digital artifacts, engaging them in authentic practices—developing models, analyzing complex data, and reasoning computationally. Within this framework, GenAI serves as both a tool and a mediator, providing adaptive, just-in-time scaffolding and enabling reflective dialogue that fosters higher-order thinking (Kasneci et al., 2023). AI is not taught as an isolated topic but is integrated as a means for students to solve real-world problems and explore complex phenomena, supporting scientific literacy, computational agency, and readiness for participation in a data-driven, AI-rich society.

Methods
Guided by design-based research, InquiryAgent integrates a Visual ML Workspace with a GenAI conversational agent. In the workspace, students use drag-and-drop blocks (Input, Preprocessing, Modeling, Testing, Explain [XAI], Visualization) to build ML workflows. The agent accesses this workspace in real-time to provide context-aware guidance, supported by a multi-agent backend (orchestrator, analyst, tutor, etc.) that uses xAPI logs for adaptive feedback. In an example activity—butterfly classification—students preprocess images, train models, and evaluate them using metrics such as accuracy and a confusion matrix, while also utilizing XAI heatmaps to interpret model decisions. Throughout this process, the agent answers questions about the domain (butterfly biology), ML concepts, or the interface, adapting its feedback based on the student's actions to promote productive struggle and authentic reasoning. (see Figures 1–2).

Results and/or Substantiated Conclusions or Warrants for Arguments/Point of View
While empirical classroom studies are forthcoming, InquiryAgent's design and theoretical basis suggest potential for supporting authentic inquiry, experimentation, and reflection with real data and models. In this paper, we present empirical data on students' user experience and learning gains in using InquiryAgent. We will also assess how the experience of InquiryAgent enhances students' interest in science learning.

Scientific or Scholarly Significance
InquiryAgent represents a novel, replicable approach for integrating interactive ML, GenAI, and XAI into secondary science education. The study extends our understanding of students' involvement in AI-based inquiry and how the experience improves their learning gains and interest in science. It contributes to the learning sciences by aligning technology with pedagogical aims—fostering agency, equity, and transparency—and provides a platform for future research on the ethical and scalable integration of AI in education.

Authors

• Matias Ignacio Rojas, University of Georgia

• Xiaoming Zhai, University of Georgia