Introducing DETECT-World: The First World Model for Deepfake Detection

Resemble AI introduces DETECT-World, a third-generation multimodal deepfake detector designed to identify manipulated audio, images, and video not only through signatures of known generators but also through learned physical-consistency checks involving lighting, geometry, motion, and audio-visual synchronization. Built on Meta AI’s V-JEPA 2 video foundation model and fine-tuned on hundreds of manipulation sources, it produces an overall manipulation probability alongside per-frame anomaly scores and spatial heatmaps. The company says adaptive training emphasizes sources the model currently misclassifies and applies real-world compression and platform-degradation augmentations equally to authentic and fake media. In internal testing, DETECT-World reportedly reached 95.8% image accuracy and 98.2% video accuracy, while its audio system achieved 99.47% accuracy on the externally validated Podonos benchmark, covering 54 languages and reducing latency to 399 milliseconds. Resemble AI reports that the model identified an unseen real-time face-swap tool with roughly 95% accuracy, though it notes that reported scores are probabilities rather than proof, long videos receive reduced-resolution analysis and no heatmaps beyond 60 seconds, and high-stakes decisions should include human review. DETECT-World is available through Resemble Detect’s API, including streaming, batch, on-premises, and air-gapped deployment options.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.