Comparing generative AI with native speakers in terms of request expressions in Japanese

Generative artificial intelligence (GenAI) tools are increasingly integrated into language education, yet their capacity to model nuanced pragmatic competence remains a critical area of investigation. This study compares the realization of Japanese request expressions by three contemporary GenAI models with those produced by 30 native Japanese speakers. Utilizing a Discourse Completion Test, responses were analyzed for formal complexity, level of directness, and the diversity of request head acts. The findings reveal significant divergences between AI-generated and native speaker outputs. Regarding formal complexity, AI models, particularly Gemini 2.0, tend to overuse address terms and display distinct patterns when employing politeness modifications. Significant differences were observed in the selection of directness strategies. While capable of generating a wide array of request forms, AI responses often lack natural spoken language features common in native Japanese, such as dialectal variations and omissions, instead adhering to a more formal, standardized register. These results suggest that current GenAI tools do not consistently replicate native speaker pragmatic norms for Japanese requests, highlighting the need for caution among learners and educators. The study underscores the ongoing gap between AI’s formal linguistic generation and its functional pragmatic competence, emphasizing the importance of critical engagement with these technologies in language learning contexts.