INTEGRATING GENERATIVE AI INTO SCIENCE EDUCATION VIA CUSTOM GPTS: CONSTRUCTION, DEPLOYMENT, AND STUDENT INTERACTIONS

Abstract

Generative AI presents new opportunities for discipline-specific, on-demand learning support, particularly in complex domains such as chemistry and physiology where students often struggle with conceptual integration and problem-solving transfer. This presentation reports on a Monash-Warwick collaborative project that develops and evaluates custom GPTs designed to scaffold student problem solving in pharmaceutical science. Using the Goldilocks Help framework (Yuriev et al., 2017) as a pedagogical foundation – which emphasises providing 'just right' scaffolding that promotes deep learning without creating dependency – we constructed course-specific GPTs that offer Socratic prompts to support metacognitive engagement, while minimising the risk of reliance and overdependence on AI-generated answers.

The implementation process involved iterative refinement, student review, and rigorous pilot testing, both for content accuracy and pedagogical effectiveness in Socratic capability. Our mixed-methods study investigates how students interact with these custom GPTs compared to open-access chatbots, drawing on usage data, perception surveys, and think-aloud interviews. Preliminary findings reveal that students adopt distinct prompting strategies depending on their familiarity with the subject matter, problem-solving mindset (instrumental versus concept-driven), and the perceived authority of the chatbot. The custom GPTs were significantly more likely to elicit structured engagement aligned with learning objectives, while open tools tended to promote less-productive trial-and-error approaches.

We discuss how the specific design features of prompt scaffolds and model behaviour can influence student cognition, affect, and self-regulation during problem solving, providing evidence-based design principles of educational AI tools. We also highlight how participatory student co-design processes contributed to the tool’s usability and perceived relevance, demonstrating the value of user-centred design in educational technology. 

As a practical resource for others looking to design customised chatbots for their courses, we offer a validated skeleton template for a system prompt and step-by-step guidance to customise it for various disciplinary contexts. This work has broader implications for the development of GenAI-driven support tools in STEM education and offers a tested, scalable model for embedding discipline-aligned pedagogical principles into generative AI applications across diverse educational contexts.

References:
Yuriev, E., Naidu, S., Schembri, L. S., & Short, J. L. (2017). Scaffolding the development of problem-solving skills in chemistry: guiding novice students out of dead ends and false starts. Chemistry Education Research and Practice, 18(3), 486-504.

ACKNOWLEDGEMENTS
This work is supported by the Monash-Warwick Alliance Education Activation Fund 2025. ChatGPT 4o was used to assist with language editing and improving readability of this abstract.