New AI system generates location-specific explanations for deepfake detection

Detecting deepfakes requires more than a yes-or-no answer. Forensic analysts need detailed explanations tied to specific parts of an image so they can verify the findings themselves. Researchers have now introduced FORGE, a system that addresses a fundamental shortcoming in how current AI models approach the problem.

The core challenge#

Standard multimodal language models—systems trained to understand both images and text—typically learn to create broad semantic summaries of entire images. This training approach, based on matching images with descriptions, optimizes the model for global understanding rather than the fine-grained, localized details that betray manipulation. As a result, these models struggle to identify the small visual cues that indicate deepfakes, such as subtle inconsistencies in a specific facial region.

Existing deepfake detection methods face another limitation: they typically specialize in either face manipulation (edited facial features) or fully synthetic content (entirely AI-generated images), but rarely handle both effectively.

How FORGE works differently#

The system tackles this inductive bias problem by routing information through two parallel visual streams. One stream comes from the multimodal model's standard encoder; the other originates from a Vision-Only Model (VOM) specifically trained on dense patch-level prediction rather than image-text alignment. Because both operate on a shared spatial grid, their output tokens can be interleaved while preserving exact positional correspondence—outperforming simpler token concatenation approaches.

To prevent the localized stream from overfitting to specific training manipulations, the researchers use a two-stage training protocol: first aligning the system on general image-caption pairs, then jointly optimizing both streams for the forensic task.

Performance and verification#

FORGE produces explanations tied to specific regions, allowing investigators to ask fine-grained questions like "Do the eyes, nose, or mouth appear real or manipulated?" Testing across both face-manipulated and fully AI-generated content shows the system substantially outperforms baseline approaches, even when evaluated on domains not seen during training. Region-specific evaluation metrics and human studies confirm that the system's explanations faithfully reflect the evidence in the image itself, rather than confabulating plausible-sounding but unfounded assessments.