Fighting Deepfake Text : Towards Building Robust defences against Arabic AI-generated Text

Abstract

The rapid advancement of Natural Language Processing (NLP), particularly through transformer-based architectures, has led to powerful large language models (LLMs) capable of generating human-like text for a variety of tasks, including question answering, content creation, and document completion. While these innovations bring transformative benefits, they also introduce ethical risks—most notably the potential for generating deceptive content at scale. Deepfake text poses a growing threat in digital ecosystems, enabling disinformation, academic fraud, and online manipulation. In response, many efforts have been made and emerged to address the challenge of detecting AI-generated text. However, existing approaches overwhelmingly focus on English, overlooking the linguistic complexity and vulnerability of Arabic. This thesis addresses that gap by proposing a novel detection framework tailored to Arabic deepfake text. Leveraging state-of-the-art transformer models and curated Arabic corpora, we develop and evaluate scalable detection techniques that account for Arabic’s richness. Our empirical results demonstrate high precision in distinguishing human-authored from machine-generated content, thereby contributing to the ethical deployment of generative AI in multilingual contexts and strengthening defences against AI-driven misinformation.