Bridging Johnstone’s Triangle with Generative AI

Generative Artificial Intelligence (AI) offers unprecedented potential for personalized Socratic tutoring in science education, yet unconstrained Large Language Models (LLMs) present profound epistemological risks. Chemistry requires students to navigate Johnstone’s Triangle—translating between macroscopic observations, submicroscopic particle kinetics, and symbolic representations. Novice learners frequently conflate these domains, assuming, for example, that physical phase changes break intramolecular bonds. Because LLMs are autoregressive engines trained on generalized, anthropocentric language, they naturally default to macroscopic analogies (e.g., suggesting the use of a "microscope" to observe molecules) when prompted to explain particulate behavior. This deterministic hallucination inadvertently reinforces students’ foundational alternative conceptions. To mitigate this risk, this paper introduces the Pedagogical Evaluation, Design, and Analysis Lab (PEDAL) framework. PEDAL is a theoretical prompt architecture that engineers strict epistemological guardrails and cognitive pacing constraints within zero-shot LLM environments. By enforcing sequential domaingating—mandating that students correctly articulate submicroscopic kinetics before attempting symbolic translation—the framework minimizes extraneous cognitive load, lowers the affective filter, and aligns conversational AI with constructivist inquiry. Furthermore, the strategic application of negative constraints explicitly prohibits the generation of macroscopic visual analogies. This framework details the theoretical necessity of structural prompting, analyzes semantic failure modes within generative examples, and discusses critical implications for science teacher education, Universal Design for Learning (UDL), Social-Emotional Learning (SEL), and the advancement of pedagogical AI literacy.