A new AI system does not just flag a photo as fake. It explains, region by region, why.
Researchers describe FORGE, a system that pairs a standard multimodal language model with a second vision model trained specifically on dense, patch-level image prediction rather than broad image-text matching. The two models process the same grid of image patches, so their outputs line up spatially - letting FORGE answer pointed questions like "Does the eyes/nose/mouth look real or fake?" with an answer tied to that exact patch of the image. The researchers argue existing multimodal models fail at this not because they lack capacity, but because their training objective rewards whole-image summaries over the small localized cues that give away a manipulation. FORGE trains in two stages, first on general image captions, then on the forensic task itself, to stop the localized vision stream from overfitting to whatever fakes it saw during training.
The more interesting claim is scope. Earlier deepfake-reasoning tools split cleanly into two camps: ones built for face-swap and face-edit detection, and ones built for fully AI-generated images. FORGE handles both, and the paper reports it holds up better than in-domain specialists when tested on manipulation types it was not trained on - the scenario that actually matters, since real-world deepfakes rarely match a lab's training set.
One footnote for the marketing department: "Forensic Reasoning with Grounded Evidence" spells F-R-G-E, not FORGE. Someone wanted the acronym more than the math.