생성형 인공지능의 지리적 문제 해결 능력에 관한 분석 연구 - 2026학년도 대학수학능력시험 9월 모의평가 세계지리 문항을 중심으로-

초록

This study explores the potential of using generative artificial intelligence (AI) as an instructional tutor in geography education by analyzing its information-search patterns and problem-solving characteristics demonstrated while solving mock-exam items without any additional prompts or task-specific guidance. To this end, the study examined the problem-solving tendencies of generative AI using the World Geography items from the September 2026 CSAT mock examination. Two state-of-the-art models—GPT-5 and Gemini 2.5 Pro—were employed, and the items were categorized into three types: stimulus-based items, map-based items, and data-based items, to identify the distinctive characteristics of each category. The quantitative analysis showed that Gemini achieved higher accuracy and faster response times than GPT, demonstrating particularly strong performance on map-based items. In contrast, both models exhibited clear limitations in interpreting data-based items, particularly in tasks requiring complex data interpretation. The qualitative analysis further indicated that the models generally recognized and reasoned accurately with raw graphical information or simple statistical data; however, their accuracy tended to decline when processing graphs derived from transformed statistical data or texts with more complex logical structures. GPT also exhibited relatively frequent hallucinations and occasionally produced errors by confusing the given stimuli with the answer choices. In addition, both models tended to infer the correct answers based on real-world data rather than relying on the contextual framework of the geography curriculum when evaluating the options. Based on these findings, this study discusses the potential and limitations of applying generative AI as an assistant teacher in geography education and presents several educational implications along with directions for future research.

키워드