Detecting Features of Artificial-Intelligence-Generated Text (A Case Study of Master's Students' Graduation Qualification Theses)
Year: 2026
Type: dissertation
Abstract: The thesis is devoted to the systematization of linguistic features indicating the use of neural language generation models in writing master's graduation qualification theses. The research material comprised twelve texts: six graduation qualification theses by master's students in Linguistics, defended between 2021 and 2025 and available in open access, as well as six analogue texts generated from the topics and tables of contents of these works by the ChatGPT, YandexGPT, and GigaChat models, together with the results of an anonymous survey of 88 students on the use of neural network technologies in writing research works. The first part of the thesis sets out the theoretical and methodological foundations of the study: it describes the basic concepts of machine text generation and the operating principles of modern neural language models and traces their evolution from n-gram systems and recurrent networks to the Transformer architecture with its attention mechanism. An overview of five families of generative models (GPT, DeepSeek, Qwen, YandexGPT, GigaChat) is given, and the choice of three priority models for the practical part is justified. Drawing on Russian and international research, the characteristic linguistic features of texts generated by artificial intelligence are systematized, existing approaches to detecting machine-generated texts are reviewed, and the specifics of academic discourse and the genre features of academic writing are described. The second part of the thesis presents a comparative analysis of six graduation qualification theses and six generated analogues. It describes the principles of sample formation, the text generation procedure, and the parameters ensuring their comparability, achieved by normalizing the quantitative indicators per 1,000 word forms. The pairwise comparison is carried out at four language levels: lexical, syntactic, stylistic, and textual. Based on the analysis, the identified features are systematized by language level and degree of diagnostic reliability; on this basis, a diagnostic set of thresholds with a decision rule for the preliminary screening of student works is proposed. The data from the anonymous student survey empirically confirm the relevance of the proposed method. The quantitative observations for each parameter are presented in the form of tables. The results of the study can be used by teachers, academic supervisors, and experts in the field of research quality control when checking student texts, including master's graduation qualification theses, in order to improve the objectivity of assessing the independence of their completion. The appendices provide a table of calculation formulas for the quantitative indicators of the comparative analysis, a script for calculating sentence-length characteristics, and a student questionnaire.
Author: Masaeva, E. D.
Language: Russian
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7212088472