The denominator behind two genuine images
A visual verification system faces errors in both directions. A fabricated image can pass as genuine; a genuine photograph can also be discredited as AI-generated. The study published on November 15 by Abdullah Yıldırmaz and Yaşar Öztürk includes 2 cases of that second kind. The remaining cases in its selected sample of 52 concern synthetic content. That detail suggests a useful requirement for detector evaluation: count both accepting generated content as real and rejecting real content as generated. Those errors can lead an editor to accept or dismiss somebody’s evidence for the wrong reason.[1]
The denominator behind that 2 needs care. The researchers did not randomly sample all images on the internet. They purposively selected 52 qualifying cases from 76 candidates documented on the verification platform between January 2024 and May 2026. Text-only material, items without demonstrated circulation in Türkiye, indeterminate cases and removed pages were excluded. This design is useful for examining the structure of disputed images. It cannot estimate the fraction of genuine photographs falsely flagged by a detector. That calculation first requires a sample representing the genuine images the system encounters.[1]
Who can reproduce the categories?
The useful contribution is to break the idea of a single fake image into different structures. Genuine and generated elements placed side by side, composite images and videos derived from photographs present different technical problems. The authors’ typology makes those combinations visible. My methodological implication is to stratify a future evaluation set by these structures. A score for an entire image may leave unclear which part a tool detected when only a small region was altered. Some observed diversity could also reflect which cases the verification platform selected, so treating these categories as the distribution of the whole internet would exceed the sample.[1]
Whether other researchers can apply the categories consistently remains open. One coder made two passes four weeks apart; the paper reports no agreement coefficient between independent coders. That does not make the typology useless. A second reading by the same person tests something different from another researcher reaching the same decision. The former examines consistency over time; the latter examines whether the categories can travel between people. Two agreeing coders could also share a mistake. Combining agreement checks with origin information verified from the image’s production process would provide a stronger evaluation.[1]
How should the next test be designed?
The four-layer schema deserves the same measurement question. Probabilistic detectors, producer-origin signals, platform labels and human examination carry different kinds of evidence. Their combination in the paper is a conceptual reconstruction from the authors’ case analysis. The study reports neither an experiment testing a new detector nor an official technical architecture disclosed by Teyit. Its value becomes more concrete within that scope: it can describe which kinds of evidence support a conclusion. Agreement between layers can enter an accuracy calculation only when their independence and comparison with verified origin are specified.[1]
A useful next step is a blinded test applying the typology to images selected from other platforms. Genuine and synthetic examples need verified origins before coders classify them without seeing those labels. False alarms and missed synthetic content should be reported with denominators for each content type. That design would examine the conditions under which mixed images can be distinguished. It should also measure the editorial review burden created when a genuine photograph is treated as suspicious. The present study contributes testable categories for building such an experiment. How those categories perform on independent examples determines the strength of the next claim.[1]