Can AI detect deepfake videos with higher accuracy than human experts in real time ?

AI systems now analyze micro-expressions, lighting inconsistencies, and biological signals to identify AI-generated faces. Platforms like Microsoft Video Authenticator can flag synthetic content before it spreads. This arms race with generative video technology is crucial for combating misinformation. Accuracy rates surpass trained investigators in controlled studies. Real-time API-based detection is already deployed.

Current AI systems can indeed detect deepfake videos with higher accuracy than human experts in controlled laboratory settings, particularly when trained on large datasets of forged and authentic media. State-of-the-art models (e.g., EfficientNet, Vision Transformers, or specialized deepfake detectors like DFDC winners) often surpass untrained human observers by identifying subtle artifacts in facial expressions, lighting inconsistencies, or unnatural blinking patterns. However, real-time deployment faces challenges: latency constraints, adversarial attacks that fool detectors, and generalization across unseen deepfake generation methods (e.g., diffusion models) remain active research areas. Some benchmarks (e.g., Deepfake Detection Challenge) show AI outperforming humans on curated datasets, but performance drops in unconstrained, real-world conditions.

— Enriched May 12, 2026 · Source: Facebook AI