New research finds that AI agents built to catch fake and manipulated photos rarely miss one - their real weakness is deciding which of their own forensic tools to believe.
A new arXiv paper dissects a training-free agentic forensics system built from specialist manipulation detectors, a triage step that screens out unreliable evidence, and a final arbitration stage that resolves conflicting reports. Researchers tested six configurations of the system across three different multimodal large language model backbones, checking performance both on familiar data and on out-of-distribution images the detectors weren't trained for. Left alone, simply fusing every detector's verdict produced a lot of false positives, flagging authentic photos as fakes. Adding triage and better prompting helped filter out bad evidence, but the single biggest driver of accuracy was the quality of the reasoning model making the final judgment call.
The paper's most striking finding is that manipulation recall - actually spotting a fake - is nearly maxed out across every setup tested. That means the bottleneck in open-world image forensics isn't detection anymore; it's calibrating how much to trust each specialized tool and resolving disagreements between them, especially when the data shifts away from what the system has seen before.
In other words, the forensic tools already know a fake when they see one. What's missing is a referee smart enough to figure out which tool is lying.