초등 수학 오개념에 대한 생성형 AI 피드백의 교육적 타당성 분석

초록

This study aims to analyze the educational validity of generative AI feedback on elementary mathematics misconceptions from the perspective of pedagogical content knowledge, and to examine where a cyclic peer evaluation among AI agents agrees with, and diverges from, the ratings of mathematics education experts. Through this, the study seeks implications for the perspectives and procedures for examining the validity of AI feedback and for the role of teachers. To this end, independent agents each produced feedback once on six cases of students’ incorrect responses from three domains of elementary mathematics, and a cyclic peer evaluation among the agents and independent ratings by three experts were conducted. The analysis showed that single-agent diagnoses contained errors, including misreadings of students’ responses, and that the peer evaluation was not consistent across agents. The expert ratings largely converged on one agent; however, even the most highly rated feedback retained a logical flaw that the peer evaluation had failed to filter out. The educational validity of AI feedback can be examined at the item level by distinguishing diagnostic accuracy from the pedagogical appropriateness of the suggestions, in light of students’ actual responses. Even when peer evaluation among AI agents is used, its judgments need to be checked against the original data, and the final judgment should rest with the teacher.

키워드