<p>Synthetic images generated by Generative Adversarial Networks (GANs) have grown alarmingly convincing, to the point where manipulated faces and scenes are often impossible to distinguish from real ones. This shift has serious consequences — from eroding digital authenticity and compromising identity security, to fueling the spread of misinformation at scale. This systematic review comprehensively surveys state-of-the-art approaches for GAN-generated image forensics, focusing on both detection and localization frameworks. Foundational architectures - including GANs, CNNs, Capsule Networks and Transformer Models - are systematically outlined, followed by analysis of dataset characteristics and distinguishing features between real and synthetic images. Detection methods are categorized into spatial-domain, frequency-domain, and hybrid multi-feature approaches, highlighting spectral fingerprinting, co-occurrence modeling, contrastive learning, anomaly-based detection, and cross-domain generalization strategies. Localization techniques that highlight tampered regions to improve interpretability are also reviewed. Comparative analysis across architectures, performance metrics, and generalization capabilities is provided, along with insights into real-world deployment limitations. Comparative analysis reveals that hybrid architectures combining spatial and frequency domain cues currently demonstrate the strongest generalization across unseen GAN models. Finally, emerging challenges — including robustness under compression, adversarial perturbations, and evolving generative models — are discussed, with promising research directions outlined toward scalable, interpretable, and generalizable GAN forensics.</p>