A new AI-detection model backs up its verdicts with visual proof instead of vague hunches.
Researchers built Defake-o3, an AI-generated image detector that zooms into suspicious regions of a photo, then checks its own reasoning against an Evidence Verifier trained on human-annotated examples. That verifier rewards the model for pointing to real, localized flaws and penalizes it for inventing artifacts that aren't there. The team also built two new datasets: GroundFake, which pairs images with bounding-box evidence and human-verified reasoning, and FakeFrontier, a tougher test set built from images made by 10 recent generators the model never trained on. In testing, Defake-o3 beat existing detectors on both accuracy and the quality of its explanations.
Most AI-image detectors today either give a bare yes/no or a text explanation that sounds plausible but points to artifacts that aren't actually there, which is useless if you need proof. A detector that can point to a specific, checkable region of a photo is a meaningfully different tool for newsrooms, platforms, and courts trying to establish provenance, not just flag suspicion. It also targets a real weak point: as image generators improve, generic explanations like the lighting looks off stop holding up, so localized, testable evidence matters more.
This is still a research paper with benchmarks it built itself, so treat claims of beating existing detectors with a healthy dose of skepticism until it's tested on images nobody involved in training it has seen.