SHALLOW ANN MODELS TO CLASSIFY UKRAINIAN AI-GENERATED TEXT

Abstract

In this study, we address the task of detecting AI-generated fragments within Ukrainian-language texts. The objective is to develop a tool capable of identifying content produced with the assistance of artificial intelligence, particularly in PDF documents related to the IT domain. The research explores and analyzes existing solutions and approaches currently available in this area. Several commercial AI-content detectors were evaluated using our custom datasets. The dataset was constructed by segmenting bachelor's theses from IT-related fields into fragments of approximately 1,000 characters each. Five artificial neural network models were tested using the custom dataset combined with a traditional NLP pipeline, achieving an accuracy of 87–88%. Given the complexity of the problem and the ethical considerations within the educational context, the classification results should be further validated by human experts. The current implementation can serve as a foundation for future improvements.