Curriculum-based GPT model development and evaluation for customized English reading material creation

This study aims to develop and evaluate a GPT-based English learning system that creates reading materials tailored to the Korean middle school first-grade English curriculum. To this end, this study adopted OpenAI’s MyGPTs platform and created the AI system through eight different developmental versions by iterating prompts and uploading relevant knowledge files. To evaluate its linguistic appropriateness, this study generated reading passages at three difficulty levels (low, medium, high) and compared them with eight textbook texts using seventeen key indices from Coh-Metrix 3. The results show that GPT-generated texts at the medium level most closely resembled actual textbook passages in sentence count, syntactic simplicity, lexical familiarity, and overall readability. Low-level outputs achieved the highest readability and concreteness scores, indicating suitability for beginner EFL learners, while high-level outputs featured greater syntactic complexity, longer sentences, and richer lexical diversity. The study also identified limitations in GPT’s consistent adherence to prescribed difficulty parameters, text-type variety, and adaptive difficulty adjustment based on learner performance.