Shallow Neural Networks for classifying Ukrainian AI-generated content

Abstract

This research addresses the challenge of detecting AI-generated content in Ukrainian IT-related documents. The study highlights the unreliability of commercial detectors on Ukrainian text and proposes a custom tool utilizing shallow Artificial Neural Network (ANN) models. Using a custom dataset of over 5,000 text fragments, the researcher developed five model architectures (M1-M5) that achieve 87-88% test accuracy. The tool provides two levels of analysis – Chunk Reports and Sentence Reports – to assist educators in identifying potential AI usage without relying on automated decision-making.