Detecting Human-to-AI Author Change in Arabic Text

Recent large language models (LLMs), such as Gemini and ChatGPT, have demonstrated the ability to produce texts that are fluent and human-like when given precise instructions, presenting significant challenges in distinguishing between human-authored and AI-generated content, particularly in morphologically rich languages like Arabic. Despite the fact that these issues have prompted several research on AI content identification, the majority of these earlier investigations framed the problem as a binary classification problem, supposing that a text is either wholly AI-generated or totally human-written. This study introduces a novel methodology for detection where the text to be identified is cowritten by generative LLMs and humans, by identifying transitions in text authorship using advanced machine learning and deep learning techniques. We propose Trans-Detect a sophisticated neural architecture that combines AraBERT with a bidirectional LSTM network and a specialized attention mechanism, specifically designed to capture subtle linguistic variations in hybrid texts. Alongside traditional Random Forest, XGBoost, and LSTM-CNN models, our detector processes Arabic text datasets to identify and predict authorship transitions. The research achieved a macro F1-score of 0.6-0.8 with traditional models, while our proposed neural architecture demonstrated superior performance with an F1-score of 0.98, showing significant improvement in detecting text origin segments and revealing the complex nature of authorship changes. The findings provide a foundation for future research exploring the integration of advanced natural language processing techniques to enhance the accuracy and robustness of style change detection systems, particularly in handling the characteristics of Arabic text.