A Systematic Review of AI-Generated Text Detection: Approaches, Tools, and Datasets

Abstract

The rapid evolution of Large Language Models (LLMs) has enabled the generation of text that is increasingly indistinguishable from human writing. While this advancement benefits various sectors, it raises significant concerns regarding academic integrity, security, and the spread of misinformation. This paper presents a comprehensive systematic review of AI-Generated Text Detection (AIGTD) techniques, evaluating their current efficacy and limitations. We categorized and analyzed various detection methodologies, including statistical and stylometric approaches, transformer-based models, watermarking strategies, and hybrid frameworks. In addition, the analysis covered 16 prominent datasets, such as HC3 and M4 for size, diversity, and limitations and domain bias, along with tools such as GPTZero, Originality.ai, and DetectGPT, which were compared on language support, usability, and detection principles. Our findings reveal that detection accuracy averages 80-99% on in-domain benchmarks but drops to 60-75% against adversarial attacks or cross-domain texts. Datasets often lack multilingual coverage and real-world diversity. Tools show high computational costs and biases toward English, with limited Arabic support. Hybrid methods outperform singles but face scalability issues. Although the field has progressed, developing robust, unbiased, and computationally efficient systems is essential. This review concludes by proposing future research directions to enhance the reliability of detection systems in an era of advancing AI.