The recent proliferation of generative artificial intelligence (generative AI), including ChatGPT, has prompted fundamental changes in teaching and learning practices across school classrooms, including science inquiry, while simultaneously revealing ...
The recent proliferation of generative artificial intelligence (generative AI), including ChatGPT, has prompted fundamental changes in teaching and learning practices across school classrooms, including science inquiry, while simultaneously revealing new educational possibilities and challenges. Generative AI can efficiently support inquiry activities by providing rich information and immediate responses. However, it also poses the risk of weakening learners’ cognitive capacities by performing core thinking processes for learners. This risk is particularly salient in science inquiry, which requires unstructured and complex problem solving. In such contexts, learners may easily offload their reasoning and judgment processes to generative AI.
Accordingly, in generative AI–based inquiry learning, enhancing learner agency—the capacity to actively lead and regulate one's own learning without relying on generative AI—has emerged as a central educational issue. However, prior studies have largely focused on identifying dependency risks, designing critical engagement systems, or examining self-directed learner characteristics. Instructional design approaches that foster learner agency in authentic classroom contexts have received limited attention. To cultivate autonomous thinking in increasingly automated environments, learner agency requires systematic teacher support and deliberate instructional design.
In response, this study aimed to develop teacher-led learner agency support strategies for generative AI–based science inquiry classrooms and to explore their effects and underlying mechanisms through implementation in real classroom settings. The research questions were as follows: (1) What learner agency support strategies are effective in generative AI–based science inquiry? (2) How are lessons incorporating learner agency support strategies implemented in practice? (3) What are the effects of learner agency support strategies in generative AI–based science inquiry? Specifically, how do these strategies influence learners’ use of generative AI, and how do they affect science interest, self-efficacy in science inquiry, and science content knowledge retention? (4) What improvements are needed for learner agency support strategies in generative AI–based science inquiry?
To address these research questions, the study employed a Design-Based Research (DBR) methodology and was conducted across two iterative cycles in collaboration with one in-service science teacher and two theoretical experts. In the first iteration, we implemented three class sessions incorporating learner agency instructional strategies with 25 eighth-grade students. In the second iteration, we applied the revised strategies using the same instructional structure with 60 eighth-grade students. Data sources included surveys on learner agency and science inquiry, content knowledge tests, generative AI prompt logs, classroom observations, and interviews. These were analyzed using mixed methods to identify effects and improvement areas.
In the first iteration, we analyzed patterns of dependent generative AI use in science inquiry. Based on prior literature and reflection with a teacher, we hypothesized that planning activities prior to AI use would enhance learner agency. Accordingly, four learner agency support strategies—(1) guiding learners to plan generative AI use, (2) helping learners understand core science inquiry concepts and procedural knowledge, (3) prompting learners to think independently before using AI, and (4) motivating learners to engage in agentic inquiry—along with eight implementation guidelines were developed and applied across three middle school science inquiry lessons. Integrated analyses revealed significant improvements in learners' overall agency, particularly in self-awareness and self-regulation within AI-based learning contexts. Prompt analyses also revealed increases in agentic interaction patterns such as critical evaluation and proactive suggestion. However, we identified limitations due to insufficient experience in prompt design, inadequate prior science concept knowledge, and time constraints. Moreover, the findings indicated that planning-based strategies alone were insufficient to secure behavioral agency in AI-based inquiry contexts that combine both generative AI technologies and the inherent difficulty of science inquiry tasks.
In the second iteration, reflections from the first iteration highlighted the need to more carefully consider differences in learners’ inquiry experience levels and generative AI competencies. Based on this, we hypothesized that learner agency in interactions with generative AI would increase as learners' foundational competencies improved, and refined the support strategies accordingly. The revised strategies consisted of four components—(1) strengthening core science inquiry concepts and procedural understanding, (2) promoting independent preliminary reflection before generative AI use, (3) systematically helping learners design better prompts, and (4) difficulty reconstructing AI-generated responses—along with eleven implementation guidelines. We applied these strategies across three middle school science inquiry lessons and comprehensively analyzed data from observations, prompt logs, surveys, tests, and interviews.
The results indicated strengthened agentic interactions, including increased use of critical evaluation prompts and proactive suggestion prompts. Survey results also showed significant improvements in self-awareness, self-regulation and selective action, and volitional control. Interviews revealed that pre-draft writing served as a reference point, suppressing AI dependency and regulating how extensively learners used generative AI. Meanwhile, prompt instruction improved learners' question regulation skills and inquiry self-efficacy. Nevertheless, we identified challenges, including difficulties in reconstructing AI-generated responses, limited prompt transfer, cognitive overload due to excessive information, and outcome-oriented dependency under task pressure. These findings suggest that subsequent designs should incorporate structured post-AI response reconstruction phases, prompt practice focused on verifying and regulating AI responses, and integration with external tools.
Synthesizing the results of the two DBR cycles, this study derived a final set of learner agency instructional strategies for generative AI–based science inquiry. The final framework consists of five support strategies and fourteen implementation guidelines: (1) strengthening core science inquiry concepts and procedural understanding; (2) promoting independent thinking and inquiry preparation before and after generative AI use; (3) supporting constructive dialogue skills to achieve inquiry goals by generating questions and reconstructing responses; (4) monitoring and promoting agentic generative AI use; and (5) designing autonomous inquiry environments and structured tools. These strategies aim to support learners in setting inquiry goals and judgment criteria independently and in critically regulating and reflecting on their interactions with generative AI, rather than relying on its immediate suggestions.
Based on these findings, the study advances discussions in three areas. First, at the theoretical level, Bandura’s concept of agency was applied to generative AI–based learning contexts, examining how agency may be weakened or transformed in environments where AI functions as an external cognitive actor. We reconceptualized learner agency from the perspective of effortful agency and articulated three conditions for its emergence in generative AI contexts: foundational competencies, independent preparation, and collaborative interaction dispositions. Second, at the practical level, the study identified foundational competencies and planning activities as critical bases for agentic participation and proposed that generative AI use competencies should be supported not merely as prompt-writing skills but as collaborative dialogue capacities encompassing question generation and response reconstruction. We also discussed how teachers serve as activity designers and learning process orchestrators who align task difficulty, time constraints, and AI response scope with inquiry phases.
The contributions of this study are threefold. First, it proposes learner agency instructional strategies for generative AI–based science inquiry and provides instructional design principles applicable to classroom contexts. Second, it extends learner agency in generative AI environments beyond tool-use proficiency to encompass effortful agency and collaborative capacities that sustain and regulate thinking. Third, by redefining teachers’ roles as designers and orchestrators of learner–AI interaction processes in AI-integrated classrooms, the study offers practical directions for mitigating concerns about cognitive delegation to AI while strengthening learner agency.