Generative adversarial networks (GANs) represent a ground-breaking advancement in artificial intelligence, revolutionizing data generation and content creation. This systematic review critically analyses literature from 2019-2024, sourced from major databases like IEEE Xplore and Scopus, to go beyond descriptive surveys and provide a synthesized, critical evaluation. We identify a pivotal gap in the existing literature: a frequent disconnect between reporting technical advancements and examining their profound ethical and societal implications. Our analysis yields several integrated insights. First, while GAN applications predominantly excel in visual content generation with architectures like StyleGAN and BigGAN achieving remarkable progress in photorealistic image and video synthesis their efficacy is highly application-specific, revealing significant trade-offs in stability and diversity in domains such as medical imaging and real-time video processing. Second, persistent core challenges like training instability and mode collapse continue to drive architectural innovations; however, we critically examine how solutions like Wasserstein loss or minibatch discrimination demonstrate varying effectiveness and limitations across different tasks and datasets. Third, and most critically, we argue that ethical concerns, particularly regarding deepfakes and data bias, are not peripheral issues but are intrinsically linked to architectural choices and the quality of training data. The key contribution of this review is its novel, integrated framework for evaluating GANs, which concurrently assesses architectural efficacy, application-specific maturity, and associated ethical risks. This holistic and critical synthesis provides researchers with a nuanced reference and outlines clear, responsible directions for future GAN development, emphasizing the need for models that are not only more powerful but also more robust, fair, and accountable.
