Abstract

Given the growing use of AI-generated materials in language education, it is essential to examine their linguistic characteristics compared with traditional human-authored texts. This study aims to compare the use of three-word lexical bundles between ChatGPT-generated reading scheme texts and human-authored Reading A-Z (RAZ) materials. Based on two comparable corpora of 403 texts each with similar token counts, three major differences were identified. First, the ChatGPT corpus uses nearly five times as many three-word bundles as the RAZ corpus. However, the RAZ corpus repeats the same bundles more often within a small number of texts. Second, the most frequent bundles in the ChatGPT corpus mainly focus on natural phenomena, whereas those in the RAZ corpus center on humans or animals. Third, the ChatGPT corpus exhibits greater structural variety in lexical bundles, including clause-related and exclusive adjective-related bundles, while the RAZ corpus favors verb- and noun-related bundles. These findings suggest that the ChatGPT corpus prioritizes linguistic variation, descriptive natural settings, and greater syntactic complexity, whereas the RAZ corpus emphasizes learnability, repetition, and child-engaging content. Pedagogically, AI-generated reading materials may require increased repetition, simplified structures, and more learner-centered topics to better support young learners’ vocabulary acquisition, reading fluency, and engagement.

This work is licensed under a Creative Commons Attribution 4.0 License.