The advent of the Fourth Industrial Revolution and Large Language Models (LLMs) has presented innovative possibilities for personalized tutoring in the field of education. However, despite technical advancements, the majority of commercialized AI tuto...
The advent of the Fourth Industrial Revolution and Large Language Models (LLMs) has presented innovative possibilities for personalized tutoring in the field of education. However, despite technical advancements, the majority of commercialized AI tutor services focus primarily on the efficiency of delivering standardized knowledge, often failing to consider the learner's cognitive context. Particularly within the exam-oriented English reading environment of Korean high schools, this component-centered approach leads to paradoxical consequences, exacerbating passive learning attitudes and causing cognitive disconnection, where learners fail to organically connect the context of the text.
Therefore, to bridge the gap between technological progress and educational practice, this study aims to integrally design the AI harnessing logic and UX/UI guidelines of AI tutors, focusing on the actual cognitive experiences learners undergo during the reading process. To achieve this, the core elements of Intelligent Tutoring Systems (ITS) were redefined from a learner-centered perspective, and a bottom-up research procedure was conducted based on data from actual educational settings rather than solely on literature review.
Through in-depth interviews with incumbent teachers, contextual inquiry of learners, and analysis of actual mobile Q&A corpora, this study identified the difficulties learners face during the reading process as five distinct types of cognitive friction. These are categorized as follows: ① Cognitive Bottleneck caused by information overload; ② Failure of Information Integration, characterized by an inability to connect contexts; ③ Fundamental Knowledge Gap, where basic schemas are absent; ④ Discrepancy Check, arising from a perceived gap between the learner's understanding and the correct answer; and ⑤ Concept Confirmation Inquiry, reflecting an active attempt to extend knowledge.
Based on these five derived types, this study established a nonlinear intervention model that responds adaptively to the learner's cognitive load state. Specifically, the logic was dualized into Group A (Types 1, 2, 3), which focuses on resolving deficits, and Group B (Types 4, 5), which focuses on expanding thought. The system was designed to adaptively switch between Direct Instruction (Mode A), which provides immediate answers, and an Inquiry-Based Approach (Mode B), which induces thought through questioning, depending on the situation. Furthermore, system prompts specifying roles and constraints were implemented to ensure this logical flow operates consistently within text-based generative AI, and the system was implemented as a 'Custom Gem' based on Gemini 3 Pro to verify the tangible validity of the logic.
Comparative experiments conducted to verify the validity of the model confirmed that the proposed system precisely diagnoses the learner's cognitive state and induces appropriate Constructive Friction. This process was found to contribute to transitioning learners from passive information receivers to active agents of meaning construction.
In conclusion, this study holds academic significance by translating the abstract concept of scaffolding in educational technology into concrete system algorithms and design language. It also possesses practical value by presenting specific UX guidelines that EdTech services should aim for in the era of generative AI. It is expected that the cognitive experience-centered design methodology proposed in this study will contribute to qualitatively innovating the interaction between AI and learners in future educational environments.