Deepfake Detection in the Era of Generative AI: A Survey and Three-Tier Reasoning Framework

Abstract

Generative AI facilitated the development of hyperrealistic deepfakes, opening creative possibilities but with serious risks for society. Recent research indicates that more than 40% of deepfakes evade state-of-the-art detectors in cross-generator tests, revealing a critical generalization gap. This survey meets that challenge by proposing a three-level reasoning paradigm for deepfake detection: (i) Signal Forensics, recording digital artifacts; (ii) Semantic Coherence, checking logical and physical plausibility; and (iii) Causal Reasoning, checking real-world plausibility. Different from existing surveys, our work (a) gives a formal treatment of causal reasoning as a new detection frontier, (b) presents a comparative study of datasets and metrics, and (c) outlines a roadmap to causality-aware, robust detection systems. By transitioning from artifact-based forensics to reasoning about reality, we contend that future detectors need to combine multimodal AI and external knowledge. This is necessary for constructing reliable digital ecosystems and countering the accelerating arms race between generative models and forensic defenses.