Oh Jieun · Park Ingyu This study analyzes the characteristics of natural language explanations generated by large language models when solving word-related items in the Korean College Scholastic Ability Test (CSAT). Word-related items from the 2022 to 2026 CSAT were input into the Gemini 3 Pro model to generate natural-language explanations, and their features were examined. The analysis indicates that these explanations exhibit two key characteristics: surface fluency and internal imperfection. This study delineates three manifestations of surface fluency: outputs relying on question components, the use of salient cues in the question, and the use of pre-trained grammatical terminology. It also identifies three manifestations of internal imperfections: the generation of multiple solution pathways, the sequential propagation of errors during the output process, and limitations in pre-trained knowledge. Because surface fluency is visible whereas internal error risk is not, grammar education that fosters users’ critical awareness is urgently required.