A generated scene can become a camera photograph without becoming a photograph of a real event. That awkward example reveals why detector errors are more varied than a simple failure to recognize AI.
Display a generated image on a screen and photograph it. The resulting file has passed through a lens and sensor. Yet the scene represented inside it still originated with a generator. Which origin should a detector report?
Before judging its answer, define the question. Whole-image generation, photographic capture, local generative editing and factual truth are different properties. A system evaluated for one may be poorly equipped to answer another.
The journey changes the available evidence
Li and colleagues' RRDataset study compares original images, transmitted versions and re-digitized versions. Re-digitization includes scanning and photographic recapture from prints, screens and projections. Their results show that robustness depends on both the detector and the route taken by the image. [1]
For DIRE, the study's Table 2 reports 89.72% of generated originals correctly detected, 90.34% after transmission and 1.42% after re-digitization. These are synthetic-class detection rates in the reported protocol, averaged across three trials. They are not overall accuracy, and the last figure is not a screen-photography-only result.
The contrast matters as much as the dramatic decline. DIRE was comparatively stable under the transmission condition. AIDE's corresponding rates were 78.95%, 74.72% and 76.04%. The study does not support the conclusion that all transformations defeat all detectors.
A screen, print and new camera pipeline can change fine detail and pixel relationships. Those changes provide a plausible mechanism for weakening some cues and introducing others. They do not isolate the cause of every individual miss. The publication reports results; an explanatory animation should not pretend to show the exact feature that disappeared.
Compression, resizing and screenshots are different tests
JPEG compression changes image information through a lossy coding process. Resizing changes spatial sampling. A screenshot may involve rendering, scaling and a new file export. Photographing a display adds an optical capture process. Treating all four as “compression” hides what the model was actually tested against.
It is equally misleading to assume that each operation always pushes the score towards photographic. Its effect can vary by model, image and processing settings. A meaningful robustness test records the sequence and compares transformed files with their own originals.
In an earlier ImageSignals research pilot, transformed images were compared with their own originals, including a simulated screenshot-like rasterization. That simulation should never be called a test of all real screenshots or physical recapture. These earlier experiments are separate from RRDataset; no pilot performance figures are used here. For the broader limits of the tools, see the methodology. [7]
False positives need photographs as carefully chosen as the fakes
With generated images defined as the positive class, a false positive is a photograph flagged as generated. A false negative is a generated image that escapes that flag. Those definitions depend on the agreed label policy; mixed-origin images need their own handling.
Why might a photograph be flagged? It may contain patterns that overlap with the model's learned distinction. Its source or processing history may differ from the photographic examples the detector learned from. These are possible mechanisms, not a diagnosis of a particular image.
It is tempting to blame dramatic light, sharpening or unusual texture after seeing an error. Unless the cause was tested, that story is speculation. A visually striking region is not evidence that the model used it.
Evaluation therefore needs a diverse photographic class as well as diverse generators. A collection of clean, familiar photographs can underestimate false alarms on the kinds of files users actually submit. Data-focused research such as B-Free addresses some of these confounds; no single construction procedure eliminates every possible one. [6]
A small generated region can be a different problem
Suppose a photograph contains one generatively replaced object. A whole-image label says little about where the change lies or how much of the file's history involves generation. Conversely, a classifier designed for fully generated pictures may not have been evaluated on that kind of edit.
The Semi-Truths project studies AI-augmented images and reports variation with editing techniques, image distributions and detector choice. Its findings resist a simple rule such as “small edits are always missed”. [2]
There is also an input-coverage issue. A preprocessing crop can omit an edited region. Semi-Truths explicitly adjusted cropping using known masks to avoid excluding the augmentation. We cannot explain its reported errors by pretending the researchers ignored that problem. On an unknown upload, the correct edit mask is usually unavailable.
This distinction changes what a useful output should say. “No whole-image generation signal detected” cannot become “no generated detail anywhere”. A location heatmap also requires validation before it can be called a map of manipulation. A visualization of model sensitivity is not automatically localization evidence.
Sometimes the ranking works and the threshold fails
A detector can place synthetic examples above photographic examples while assigning the wrong labels at a fixed threshold. Ranking and decision quality are different measurements.
Epstein and colleagues provide a concrete example in their chronological generator evaluation: the GLIDE condition had perfect AUC but only 15% synthetic-class accuracy at the operating rule discussed in the paper. That is a result for a particular experimental condition, not a statement about every GLIDE image or present-day service. [3]
Imagine two groups of illustrative scores. Every photographic score lies below every synthetic score, but the deployed threshold lies above most of both. The ordering separates the groups; the threshold misses many synthetics. Moving it could help on those data.
That remedy has limits. A threshold cannot recover a region excluded by preprocessing. It cannot recreate information lost in a capture process. And selecting a threshold by looking at the final test labels would overstate independent performance. Threshold development and evaluation need separate evidence.
Confidence does not establish applicability
Calibration concerns reliability across groups of predictions. A model can be well calibrated on familiar validation data and become less reliable after the input distribution changes. Ovadia and colleagues demonstrate this broader uncertainty problem in classification experiments; their work does not supply a false-positive rate for ImageSignals. [4]
A single confident mistake does not prove miscalibration. Even a reliable high-probability group permits occasional errors. The question is how predictions behave across comparable images with independently established origins.
Base rates introduce another distinction. Consider a hypothetical set of 10,000 images containing 100 generated examples. At 95% generated-image recall and a 2% photographic false-positive rate, the expected counts are 95 true detections and 198 false alarms. Only about 32.4% of flags would identify generated images.
Those are invented assumptions for transparent arithmetic, not detector measurements. They show why “few photographs falsely flagged” and “most flags are correct” are not equivalent claims. Precision depends on prevalence as well as the operating characteristics.
Progress has to be tested on the right journeys
The limitations do not make detection pointless. Community Forensics reports improved transfer through broader generator coverage. That is encouraging evidence within its evaluations. It does not remove the need to test photographic sources, transmission, recapture and local editing separately. [5]
When examining a result, record what is known about the file's journey. Is it an original export, a repost, a screenshot or a photograph of a display? Does the task include local edits? Was this detector tested under similar conditions? If the answers are unknown, preserve them as unknowns.
Tools can support that inquiry. ImageSignals' region search locates possible matches for a named object within the uploaded image; it does not retrieve earlier copies from the web. Object detection and image description can help organize visible content, but they do not certify origin. Their value here is to assist inspection and form better questions. Finding earlier publications requires separate source research.
The useful conclusion is a bounded one
A wrong answer may reflect missing evidence, an unfamiliar image distribution, a poor operating threshold or a mismatch between the task and the label. These failures need different responses. Calling all of them “the detector was fooled” hides the distinction.
Return to the photograph of the generated scene. The camera was real. The depicted scene was invented. A useful evaluation must say which fact the detector was expected to identify—and demonstrate that it can do so on files that have taken that journey.
Try the AI image detector, then interpret its output alongside the available history and methodology.
Watch the video
Why AI Detectors Get It Wrong · ImageSignals · 4:56.
This article draws on research developed for the ImageSignals video series and expands it for readers.
Sources & further reading
- Li et al., Bridging the Gap Between Ideal and Real-world Evaluation, ICCV 2025, §3.3.2 and Table 2.
- Pal et al., Semi-Truths, NeurIPS 2024 Datasets and Benchmarks, experiments and Appendix C.1.
- Epstein et al., Online Detection of AI-Generated Images, ICCV Workshops 2023, §3.3 and Figure 2.
- Ovadia et al., Can You Trust Your Model's Uncertainty?, NeurIPS 2019.
- Park & Owens, Community Forensics, CVPR 2025, §4.3.
- Guillaro et al., B-Free, CVPR 2025.
- ImageSignals, Methodology and transparency. The earlier pilot described here is separate from the cited studies; no experiment was rerun for this article.