A March 2026 peer-reviewed systematic review (34 studies, 2014–2025, PRISMA methodology) from Sol Plaatje University finds that deepfake detection research has structurally shifted from CNN-based pixel-artifact methods toward transformer- and CLIP-based architectures capable of cross-modal semantic analysis, tracking the increasing realism of generative AI outputs. Persistent gaps identified include multimodal cross-dataset generalization failures, explainability–robustness trade-offs, and a substantial disconnect between benchmark performance and real-time operational deployment. For financial institutions, the review underscores that detection tooling for synthetic media fraud (voice, video, document) remains technically immature for production environments, elevating near-term operational and compliance risk from AI-generated deception.
145 other tracked developments touching the same claims.