UCLA researchers built an AI system that uses light-based analysis to examine more than a dozen videos simultaneously, detecting deepfakes with nearly 98% accuracy, announced October 1, 2026. The approach analyzes subtle physical light-interaction patterns that are difficult for current video-generation models to replicate convincingly.
Deepfake detection has become an arms race problem: each generation of detection technique eventually gets incorporated into the training data of the next generation of generative video models, which then learn to avoid the specific artifacts the detector was looking for. A 98% accuracy figure is a meaningful result today, but the more important question is how long that number holds up once generative models are trained against it specifically.