<![CDATA[This study compares the assessment of Japanese sakubun by a native Japanese teacher and two large language model based AI tools, Gemini AI (3.1) and ChatGPT (5.5), and examines the extent to which AI generated scores align with human assessment. The data comprised eight sakubun texts on the theme of meiwaku (inconvenience) written by second year students with approximately JLPT N4 level Japanese proficiency. Each sakubun was assessed using the same rubric covering four dimensions: content, organization, expression, and accuracy, as well as an overall score. The scores were descriptively compared across the three evaluators at both the criterion and overall levels. The results showed that the Japanese teacher assigned the highest average overall score of 16.25, followed by Gemini AI with 14.25 and ChatGPT with 13.75. Of the two AI tools, Gemini AI demonstrated closer overall alignment with the teacher’s assessment. Both AI systems also exhibited a score compression tendency. Higher quality sakubuns were generally evaluated lower than by the teacher, whereas some lower quality sakubuns received relatively higher scores. The most substantial discrepancies were observed in content and expression, indicating that AI assessment was less sensitive to idea development, contextual appropriateness, naturalness of expression, and subtle linguistic nuances. These findings suggest that although Gemini AI and ChatGPT can provide useful supplementary assessment, human evaluator judgment remains essential, particularly for evaluating qualitative and context sensitive aspects of Japanese writing. ]]>
