Comparing ChatGPT and Authentic Discourse as L2 Academic Speaking Partners: A Corpus-based Analysis and Exploration of Pedagogical Applicability

This study examines the extent to which spoken academic English produced by the ChatGPT-4o model (ChatGPT) aligns with authentic discourse and is applicable for use in L2 speaking materials. As generative AI tools gain popularity among L2 learners for speaking practice, it is essential to assess the linguistic and discourse-level validity of such output. Ten dialogues generated by ChatGPT were compared with an authentic transcript of a doctoral dissertation defence from the Michigan Corpus of Academic Spoken English (MICASE) using a corpus linguistic and content analysis approach. Quantitative analyses revealed that ChatGPT produced language with higher lexical diversity, greater use of advanced vocabulary, and more complex syntax than the MICASE sample, but with lower readability. Content analysis showed that while ChatGPT simulated turn-taking and appropriate question–answer sequences, it lacked features of spontaneous interaction such as clarification requests, topic shifts, and overlapping speech. Speaker roles in the generated texts were consistent but followed more scripted and idealized patterns. These findings suggest that ChatGPT can serve as a useful supplementary tool for academic speaking practice, especially in providing structured, high-level input. However, its limitations in simulating authentic interaction highlight the need for guided integration alongside real human dialogue. The study is expected to contribute to the growing body of research on AI-assisted language learning.