Academic vocabulary use in research publications in the AI era: human-authored vs AI-mediated texts

Abstract

This study addresses an underexplored dimension of computer-assisted language learning (CALL) by developing the ChatGPT-Mediated Academic Word List for Research Publication (CAWL-RP), an exploratory, corpus-specific lexical profile rather than a general-purpose academic word list, and by examining how AI-mediated rewriting may recalibrate academic vocabulary within one controlled corpus design. Drawing on a 16.9-million-token corpus, the data were systematically processed through de-authorship, generalisation, and re- academisation to simulate AI-supported research writing. Subsequently, a set of systematic procedures comprising frequency, range, uniformity, dispersion, lexical profiling, a meaning-based approach, and concordance line analysis was then employed to construct the CAWL-RP, which comprises 869 word families and 2,341 types. Comparative analyses with the Academic Word List, the New Academic Word List, and the Academic Vocabulary List reveal both convergence and divergence. While the CAWL-RP shares a substantial lexical core with established lists, 470 items occur exclusively in the AI-mediated corpus. Moreover, the CAWL-RP is characterised by a markedly higher proportion of general, high- frequency academic vocabulary, whereas the comparison lists contain relatively more low-frequency and field-specific items. Rather than merely documenting lexical variation, the study reconceptualises academic vocabulary as a dynamic construct that may be partially reconfigured under AI-mediated conditions. By situating the CAWL-RP within CALL and English for Academic Purposes theory, the findings empirically demonstrate how generative AI may operationalise academic lexis differently from human-authored corpora. Because the CAWL-RP 1is derived from a constructed AI-mediated rewriting corpus based on human- authored 2020 source texts and refined under a single model (GPT-3.5) and a fixed prompt-and-session protocol, it is best interpreted as a provisional lexical profile of this specific corpus design rather than a generalisable inventory comparable in status to the AWL, NAWL, or AVL. Within this bounded scope, the study contributes to ongoing debates on how academic vocabulary should be defined, taught, and operationalised in AI-supported learning environments.