JGAP 認証申請における資料作成支援を目的とした生成AI の出力特性の実証分析

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

As JGAP certification has become more widespread, the administrative burden of preparing risk assessment documentation has significantly increased. Consequently, the application of generative AI in the agricultural sector has garnered increasing attention. However, empirical research evaluating its practical utility in JGAP-compliant document generation remains sparse. This study aims to clarify how generative AI can facilitate the identification of hazards and the formulation of management measures, while simultaneously identifying the requisite conditions and constraints for its practical implementation. To evaluate these capabilities, the output characteristics of generative AI in risk assessment documentation were analyzed across three dimensions: stability, lexical features, and information quality. Using a tea farm in Shizuoka Prefecture as a case study, 24 prompt patterns were developed by varying the volume of input information and specific instructional constraints. A total of 240 documents were subsequently generated using ChatGPT (GPT-4o) and Gemini 1.5 Pro. The results showed no statistically significant differences in output length stability across prompt conditions or between models. In contrast, lexical composition varied depending on the information provided in the prompts. Furthermore, differences between models were observed in information quality evaluations (accuracy, completeness, novelty, and format) conducted by a JGAP internal auditor. These findings suggest that generative AI has the potential to support the identification and examination of hazards and management measures. Therefore, generative AI should be positioned not as an autonomous document-generation agent, but as a decision-support tool that presupposes human judgment.