Contents
Real photographs get flagged as AI, and it happens most after ordinary processing: a crop, a re-save, an upload that compresses the file. The reason is mechanical. A detector keys on a faint production fingerprint, and real-world capture and compression can either erase the fingerprint of a real photo or add noise that a detector misreads as the fingerprint of a fake. A false positive is not the tool malfunctioning. It is the tool doing exactly what it was trained to do, on an image outside what it was trained on.
The scale of the problem
Independent testing shows false positives are common, not rare. Bellingcat, testing the detector “AI or Not” on twenty genuine photojournalism-contest photographs, found it wrongly flagged 6 of 20 as AI-generated. That is a real photograph called fake almost a third of the time, on exactly the kind of image where a wrong answer does the most damage.
The rate also depends heavily on which tool you use. NewsGuard tested five detectors on the same 45 images and recorded very different false-positive behavior:
| Detector | Real images wrongly flagged as AI |
|---|---|
| ScamAI | 40% |
| ZeroGPT | 20% |
| AI or Not | 6.67% |
Two of the five tools in that test flagged none of the images, while the worst flagged 40 percent. The tool you happen to pick changes your odds of being wrongly accused more than anything about the image itself.
Why a real photo trips the wire
The detector is reading a signature, and the signature is not unique to AI. Frank, Eisenhofer, Schönherr et al. (ICML 2020) traced the frequency-domain trace that many detectors read to upsampling, a step in how generators build an image; the full mechanism is in how AI image detectors actually work. But real-world pipelines also resample, sharpen, and compress, and those operations leave periodic patterns of their own. When a real photo’s processing happens to resemble the generation trace, the detector fires.
The deeper problem is that detectors often learn the wrong thing outright. Grommelt, Weiss, Pfreundt et al. (2024) showed that standard detection datasets are biased by JPEG compression and image size, and that detectors “indeed learn from these undesired factors.” If the real images a detector trained on were compressed a certain way and the generated ones were not, the detector partly learns “this compression means real” or the reverse, rather than anything about generation. Feed it a genuine photo compressed the unfamiliar way, and it confidently calls it fake. That is a false positive baked in at training time, and it explains why re-saved and downsized real images are where false alarms cluster.
Novel content makes it worse. A detector flags what it does not recognize, so genuine images from an unusual camera, in an unusual style, or of unusual subjects can read as anomalous. The same fragility that lets a detector generalize poorly to a new generator, documented for feature-space detectors by Ojha, Li, Lee (CVPR 2023), also lets it stumble on real images it has no reference for.
Reading a false positive correctly
Because the mechanism is a fingerprint that real processing can imitate, a single high score against your own genuine image is weak evidence. Wang, Wang, Zhang et al. (CVPR 2020) built one of the standard pixel-based detectors, and even well-built detectors of that kind inherit the training biases above. So a confident “AI” verdict on a photo you know is real is not a paradox to explain away. It is the expected error rate showing up.
The practical response is the same as for any detector output. Test the original file rather than a re-compressed copy, because the extra compression is often what pushes a real image over the line, a failure mode covered in can AI images be detected after compression. Get a second opinion from a tool that reads a different signal, since the spread in the table above means one tool’s false positive is often another tool’s correct pass. And weigh evidence off the file, such as the original source or the photographer’s own record, which outranks any score.
For creators, this cuts two ways. A false positive can put honest work under suspicion, and there is no perfect defense, only the record you keep and the ability to show the original capture. The tools that aim to protect real artwork from being scraped into training sets are a separate line of defense and do not change how a detector reads a finished file. What does change your position is understanding that the flag is a probability from a narrow reader, not a judgment. For how far these tools can be trusted overall, see are AI image detectors reliable.
Sources
- Bellingcat (2023). Testing ‘AI or Not’: How Well Does an AI Image Detector Do Its Job?
- NewsGuard (2024). Leading AI Image Detection Tools Mislead Online Users, Often Declaring Authentic Content Fake.
- Grommelt, Weiss, Pfreundt et al. (2024). Fake or JPEG? Revealing Common Biases in Generated Image Detection Datasets.
- Frank, Eisenhofer, Schönherr et al. (2020). Leveraging Frequency Analysis for Deep Fake Image Recognition. ICML 2020.
- Wang, Wang, Zhang et al. (2020). CNN-generated images are surprisingly easy to spot… for now. CVPR 2020.
- Ojha, Li, Lee (2023). Towards Universal Fake Image Detectors that Generalize Across Generative Models. CVPR 2023.