불교를 바라보는 인공지능의 시선 - 거대언어모델(LLM)의 불교 지식 비교 분석 -

초록

This study aims to empirically examine how accurately and reliably large language models (LLMs) can understand and reproduce Buddhism as a specialized and culturally embedded system of knowledge. To this end, a dataset of 90 evaluation items was constructed based on The Predicted Question Book for the Buddhist Missionary Examination published by the Buddhist Promotion Bureau of the Jogye Order of Korean Buddhism. Reflecting the actual structure of the missionary examination, the dataset was organized into two sessions, each consisting of 45 questions (40 multiple-choice items and 5 open-ended questions). The models evaluated in this study include OpenAI GPT-5.2, Google Gemini 2.5 Pro, Anthropic Claude Sonnet 4.5, and DeepSeek Chat. All experiments were conducted via APIs with the temperature parameter fixed at 0, and each model was required to produce structured JSON outputs containing the selected answer, the reasoning process, and a self-reported confidence score. The results show that the Gemini and Claude models achieved high accuracy rates, but exhibited limitations in confidence calibration. The OpenAI model demonstrated a relatively stable distribution of confidence levels, whereas the DeepSeek model displayed a pronounced risk of high-confidence errors, showing low accuracy despite consistently high confidence scores. These findings indicate that while LLMs perform strongly at the level of knowledge (知, ji)—that is, the conceptual explanation of Buddhist doctrines—they exhibit structural limitations at the level of wisdom (慧, hye), which is shaped through ritual practice, transmission, and embodied experience. This study thus empirically reaffirms the classical Buddhist epistemological distinction between knowledge and wisdom in the context of contemporary AI systems, and argues for a reconceptualization of artificial wisdom not as wisdom inherent to AI itself, but as an instrumental infrastructure that supports human practice and reflective cultivation.

키워드