Investigation Into the Identification of AI-Generated Short Dialectal Arabic Texts

Identifying AI-generated text (AIGT) has become crucial in addressing the potential misuse of generative AI tools and their impact on trust, fairness, and authenticity of content. This study investigates the detection of AIGT in short dialectal Arabic texts, an area that, to our knowledge, has not been previously explored. Two datasets were constructed featuring texts that were rephrased or generated to match the topics and length of MADAR CORPUS, encompassing dialectical human-written texts (HWTs) originating from five major Arabic cities. Among several AIGT detection methods, the pre-trained language models AraELECTRA and AraBERT outperformed both machine learning and feature-based approaches. Moreover, we analyzed the key features that significantly influenced the accuracy of the model.