AI Image Detector False Positives on Human Art: Every Published Number (2026)

1 How Often Do Image Detectors Flag Real Pictures as Fake?

The spread is the point. A range that wide is not noise and it is not a difference between good tools and bad tools, because it is the same tools being measured twice. It is the difference between two defensible ways of deciding where likely AI begins. A vendor picking the lower figure is not fabricating anything. It is reporting the calibration that flatters its product, in a market where no convention forces it to report the other one.

For anyone on the receiving end of a flag, this collapses the distinction between a confident verdict and a hedge. A detector reporting high confidence is reporting confidence relative to a boundary somebody drew, and the paper that produced these figures found that redrawing the boundary against current imagery pushed several detectors past the point where most of the real photographs they saw came back as fake. The same structural problem shows up wherever detection is used as evidence, which is why institutional rules on using detector output as proof have been tightening rather than loosening.

It is also worth being precise about what the wide figure is and is not. It is not a claim that detectors flag most pictures most of the time in ordinary use, because most deployments run on the gentler calibration. It is a claim about how little the reassuring figure constrains real-world behaviour: a rate quoted from one calibration tells you almost nothing about performance under another, and buyers are never told which one they are getting. Treat any single published rate as a description of a test, not a promise about your file.

2 Why One Detector Reports Two Different False Positive Rates

Raising the cutoff buys a lower false positive rate and pays for it in missed AI images. Lowering it does the reverse. There is no setting that gives both, and the research is explicit that none of the detectors evaluated managed to beat chance on modern AI content while keeping a defensible rate of false alarms on real ones. That is not a tuning problem waiting to be solved. It is the shape of the trade-off at the current state of the art.

The mechanism behind it matters for artists specifically. Modern phones do not simply record light; their processing pipelines fuse multiple sensor reads and suppress noise and motion blur, so the device generates the photograph as much as it captures it. Detectors read that generated quality as synthesis. Digital art tooling performs the same class of processing through brush engines, filters and upscalers, which makes the extension plausible, though it has not been measured on artwork and this page does not claim it has. Readers who want the text-side version of the same argument will find it in the work on how detection accuracy collapses outside the conditions a model was tuned for.

3 What Is Actually Inside a Vendor's Art Benchmark?

A photographed oil painting and a finished digital illustration are different objects to a pixel classifier. One arrives through a camera pointed at canvas, carrying brushwork, craquelure and the texture of a physical surface. The other is born in software, built from layers and synthetic brushes, exported through a compression pipeline. A benchmark built entirely from the first tells you very little about the second, and the artists reporting problems are almost all working in the second.

The second vendor figure has the same shape. It covers pre-2022 web images, described as photographs, web graphics and computer-generated imagery, which again excludes the category under dispute. None of this makes the numbers false. It makes them answers to a question nobody was asking. The site’s own reading of that vendor’s text detector, in a review that found its near-zero false positive claim held up better than its false negatives did, reached a similar conclusion by a different route: the headline number is usually real and usually about something narrower than it sounds.

There is one more twist worth recording. The same vendor also ran an evaluation against images drawn from a subreddit where artists gather to ask whether a picture is AI, and reports classifying every human sample in it correctly. That community is simultaneously one of the places where false positive complaints are accumulating fastest. Both things can be true, and the likeliest explanation is again the sample: images that reached a clear community consensus are the easy cases, and the disputed uploads that generate the complaints are precisely the ones a consensus filter removes.

4 How Big Are the Independent Image Detector Benchmarks?

A comparison run on a handful of images cannot produce a false positive rate in any meaningful sense, and to their credit neither of these publishes one. What they publish is a ranking, and rankings from samples that small are close to arbitrary; swapping a single image can reorder the table. Those rankings are nonetheless what most roundups, buying guides and forum recommendations are ultimately built on, several citations downstream from the original.

This is the real asymmetry in the market. The vendor benchmarks are large and self-administered. The independent benchmarks are honest and tiny. The only large independent evaluation is academic, measures photographs rather than art, and reaches a conclusion none of the buying guides repeat. Anyone deciding whether to trust a verdict is choosing between those three kinds of evidence, usually without being told which one they are looking at, much as students choosing a pre-submission checker are rarely told how thin the evidence behind most detector accuracy claims actually is.

5 What a Flag Actually Tells You About Your Artwork

There is also a second detection channel that gets conflated with the first and works nothing like it. Platform labels such as the made-with-AI badges applied by large social networks read provenance metadata embedded in a file, while detectors classify pixels. Metadata travels with a file and can be inherited from a brush pack, a filter or an editing tool, which means a wholly handmade image can carry a signature it never earned. It can also be stripped, which is why the people generating images spend their time removing it. The two systems fail in opposite directions, and the same confusion surrounds the watermark checkers that claim to read signals that are not in the file.

Publishing platforms, galleries and marketplaces increasingly act on these verdicts, which raises the cost of the error well beyond a bruised ego. An artist who cannot clear a flag can lose a commission, a listing or an account, and the appeal process usually involves proving a negative to someone holding a number they believe is objective. The institutions furthest along this road have started backing away from that posture, which is the context behind the universities that switched their detectors off rather than keep adjudicating with them.

The practical reading is unglamorous and reasonably firm. A single verdict from a single detector is weak evidence about any individual image, whatever confidence percentage sits next to it. Agreement across several detectors is worth more, though less than it looks, because detectors share training data and tend to share mistakes. What holds up is process evidence: layered working files, version history, timestamped drafts, the messy intermediate states that a generated image does not have. That is the same advice that survives on the text side, where the first move after an accusation is to assemble the record rather than argue with the tool, and where the groups who lose most from detector error are consistently the ones whose ordinary output sits closest to whatever the model calls machine-like.