LEXICAL-SYNTACTIC AND DISCOURSE FEATURES OF AI-GENERATED SCIENTIFIC TEXTS

Abstract

Generative language models are becoming increasingly convincing, and distinguishing between “machine” and “human” academic style is becoming challenging – especially in Russian-language sociology. The authors compare 15 man-made annotations and 45 annotations, generated by three neural networks on the same topics. The goal was to find lexical, syntactic, and discourse markers that distinguish human text from model text. The method combines qualitative analysis, frequency counting, the χ² test for categorical features, and descriptive statistics for quantitative features. The result: AI-generated texts more frequently use transitional and frame formulae, more impersonal constructions, a more rigid composition, less variability in sentence length, and lower TTR and MTLD scores. According to our data, it is impossible to reliably identify “machine” text based on a single feature – only a combination of markers works. Practical value – in teaching academic writing, editorial and expert practice in evaluating texts created with the participation of AI.