Deepfake detection is losing ground to generative models

This article critiques the current state of deepfake detection, arguing that core forensic assumptions (visible composites, model fingerprints, temporal inconsistencies, physiological cues, and signal survivability after re-encoding) are failing as generative models improve. The authors propose augmenting media forensics with a communication-layer analysis—assessing authority fit, conversational coherence, and pressure tactics—and emphasize tried-and-true procedural controls (callback verification, out-of-band confirmation, challenge questions, separation of authorization channels) as the most reliable defenses observed in documented deepfake frauds.