VERIFYING THE PRESENCE OF AI-GENERATED TEXT IN ARABIC AND ENGLISH WRITINGS

Abstract

The accelerating progress of generative AI has led to growing concerns about the authenticity and source of text content, particularly in academic and professional contexts. While several detection systems already exist, a significant challenge remains in developing frameworks that achieve high detection accuracy, remain robust across varying text lengths, and support multiple languages, including the Arabic language. This study proposes a detection framework, the Hybrid Fusion System (HFS), for identifying AI-generated text. The proposed framework is an ensemble system that contains two complementary components: the Feature Engineered Classifier (FEC) and the Language Model Classifier (LMC). The first component uses two type of features for AI-generated text classification: the linguistic features extracted from the input text, and probability features generated by combining token probabilities obtained from a series of language models using scalar and vector functions. The second component includes fine-tuned language models adapted for the text classification task. Finally, the outputs of the FEC and LMC are combined using an ensemble mechanism that synthesizes their outputs to produce the final classification result. In the first stage, our proposed system was designed to recognize English-language text written by AI, and then adapted to support the Arabic language through necessary linguistic and structural modifications. The study is also introduces a novel Arabic and English abstract dataset. This dataset is curated specifically for detecting AI-generated abstracts submitted by students at An-Najah National University. The proposed system achieved 99.76% accuracy on the English dataset and 99.55% accuracy on the Arabic dataset. Furthermore, it demonstrated substantially stronger performance than three publicly available AI-text detectors: ZeroGPT, Sapling, and Detecting- AI. On the English dataset, these tools reached only 94–96% accuracy, while on the Arabic dataset their performance dropped sharply to 56–73%.