Linguistic Fingerprints of AI-Mediated Academic English: A Corpus-Based Study of Lexical Density, Stance, and Authorial Voice

Abstract

The rapid normalization of generative artificial intelligence in academic writing has created a linguistic ecology in which academic texts are increasingly not simply human-authored or machine-generated, but AI-mediated. This study investigates the linguistic fingerprints of AI-mediated academic English through a corpus-based comparison of lexical density, stance, and authorial voice. The proposed corpus consists of three balanced subcorpora: human-written academic texts, AI-generated academic texts, and AI-mediated texts revised by human writers after documented AI assistance.

Drawing on systemic functional linguistics, metadiscourse theory, corpus stylistics, and contemporary research on human-AI writing, the study measures lexical density, lexical diversity, nominalization, hedging, boosters, attitude markers, engagement markers, self-mentions, and voice distinctiveness. The quantitative model includes descriptive statistics, reliability testing, normality testing, ANOVA or Kruskal-Wallis tests, effect sizes, correlation analysis, principal component analysis, cluster analysis, and supervised classification. The qualitative component interprets concordance lines and rhetorical patterns.

The central argument is that AI-mediated academic English is not identifiable through a single marker; rather, it is characterized by a patterned convergence of lexical polish, controlled stance, reduced authorial risk, and rhetorically smooth but sometimes pragmatically flattened voice. The study contributes a replicable framework for evaluating AI mediation ethically without reducing academic quality to plagiarism detection or mechanical authorship attribution.