The Vanishing of Human Touch Through Mechanical Tone in AI Polished Academic Writing

Abstract

ABSTRACT
The rapid adoption of large language models (LLMs) in scholarly writing has been documented largely through lexical markers, such as the sudden rise of words like "delve" and "intricate" in published abstracts. Less understood is what disappears alongside this vocabulary shift: irregular sentence rhythm, first-person hedging, and the small idiosyncrasies that once signaled individual authorial voice. This study examines whether AI-assisted polishing produces a detectable flattening of "human touch" in academic prose, operationalized through sentence-length variability, hedging density, first-person presence, and lexical idiosyncrasy. Using a qualitative textual analysis and stylistics framework, this study compared eleven paired academic paragraphs across ten disciplines, each written in an unassisted human register and then rewritten in a generic AI-polished register. Both sets were then subjected to computational stylistic measurement. Results showed markedly reduced sentence-length variability, near-total absence of first-person markers, and a concentration of AI-associated lexical items in the polished set. These findings, though drawn from a small illustrative corpus, align with broader evidence of stylistic homogenization under LLM assistance, raising pedagogical and disciplinary questions about equating polish with improvement.