The Imitation Game revisited: A comprehensive survey on recent advances in AI-generated text detection

In recent years, AI-generated text detection (AIGTD) has attracted more and more attention, with numerous novel methodologies being proposed. However, most existing reviews on this topic tend to be fragmented and incoherent in content, lacking a coherent and comprehensive framework for understanding. This paper comprehensively analyzes and summarizes the latest advancements and prominent technologies in this fastmoving field. In order to do that, we introduce a novel comprehensive multi-level taxonomy for AIGTD approaches, where the existing research can be broadly categorized into three directions, tackling the key challenges of classifier training, intrinsic attributes, and information embedding, respectively. To help researchers and practitioners understand and address detection and attack scenarios, we also introduce a classification of black-box and white-box models based on interpretability and transparency, as well as the computational requirements required to use the baseline methods. Moreover, we carefully provide a comprehensive performance comparison and analysis across several datasets for these methods, collect commonly used benchmark datasets, and outline potential future research directions in this field. To facilitate sharing, we consistently maintain the relevant materials at: https://github.com/Nicozwy/AIGTD-Survey.