Linguistic Features of AI-Generated Academic Texts and the Role of Human Editing

Abstract

The rapid spread of large language models (LLMs) has significantly transformed academic writing practices and actualized discussions about authorship, language quality, and academic integrity. At the same time, diachronic changes in academic discourse during the active implementation of generative artificial intelligence remain insufficiently studied. The study combines a systematic literature review with a corpus diachronic analysis of authentic academic annotations, allowing us to trace long-term trends in the development of academic discourse. The aim of the work is to identify linguistic changes in academic writing during 2015–2025 and determine the role of human editing in quality assurance of AI-assisted scientific texts. The research material was a corpus of 870 English-language annotations of scientific articles indexed in the Scopus database in the field of arts and humanities. Quantitative linguistic analysis was carried out using Python tools. Indicators of lexical density, lexical diversity, syntactic complexity, average sentence length and frequency of cohesive markers were analyzed. A statistically significant increase in lexical density and frequency of cohesive markers has been revealed, indicating an increase in information compression and explicit discursive organization of texts. Indicators of traditional lexical diversity and syntactic complexity remained relatively stable. The observed trends are consistent with characteristics described in AI-assisted writing studies; however, the study design does not allow them to be directly related to the use of large language models. Human editing remains a key factor in ensuring factual accuracy, discursive coherence, lexical enrichment, and academic integrity. The results can be used to develop practices for the responsible use of generative AI in academic communication.