The rapid development of generative artificial intelligence has enabled the creation of highly realistic synthetic images and videos. Although these technologies support beneficial applications in entertainment, education, accessibility, and digital content production, they also enable impersonation, misinformation, fraud, and non-consensual synthetic media. Reliable detection of AI-generated deepfake content is therefore an important problem in computer vision, multimedia forensics, and information security. This paper reviews the principal techniques used to detect manipulated images and videos, including spatial artifact analysis, frequency-domain analysis, physiological signals, temporal inconsistency modeling, multimodal verification, and provenance-based methods. It also examines representative datasets, evaluation metrics, generalization problems, adversarial attacks, and emerging detection architectures. The analysis indicates that current detectors can achieve high performance on familiar datasets but often degrade substantially when evaluated on unseen generators, compression levels, identities, and post-processing operations. Future systems should combine visual analysis with temporal reasoning, provenance metadata, robust uncertainty estimation, and continuous adaptation to new generative models.