Yes. The same image can pass two AI detectors and still be flagged by another. A detector score describes one system's response to one file. It is not a shared measure that every other detector must agree with.
The distinction matters if an image humanizer advertises a low score. You need to know whether the tool was tuned using that detector and whether an independent detector checked the finished output.
What each detector is answering
Illustration of how to read different measurement types. No synthetic detector scores are shown.
Watermark verifier
What this means
This asks about one embedded signal, not whether the picture was generated by every possible model.
Visual detector A
What this means
Its score is meaningful within its documented system. Do not average it with another vendor’s score.
Visual detector B
What this means
Disagreement is a finding to investigate, not a reason to hide the less favorable result.
A detector used for tuning is not an independent judge
Imagine editing a picture while repeatedly asking the same detector whether it passes. That can teach the editing process to satisfy that detector's particular rules. A different detector may respond differently.
Deprint tested this directly on two scenes. Small changes were first optimized against a local detector called NPR. A follow-up also optimized against a second local model, SigLIP/DINO. Both local scores fell below their thresholds. Hive, which was not used for optimization, still flagged the exported files.
| Final file | NPR, used during optimization | SigLIP/DINO, also optimized | Independent Hive check |
|---|---|---|---|
| Desk | Below threshold | 0.1500%, below threshold | 99.9%, positive |
| Bakery | Below threshold | 0.0119%, below threshold | 99.9%, positive |
The reported percentages belong to different models. Do not compare their sizes as if all three measured the same probability. The useful comparison is their decision on each exact file.
A tiny final edit can hide a larger earlier change
The follow-up edits were bounded around images that had already gone through DeSynth reconstruction. Describing the entire workflow as a tiny or invisible edit would leave out that earlier processing.
| Scene | Similarity of final file to its already-processed input | Similarity of final file to the original |
|---|---|---|
| Desk | SSIM 0.9816 | SSIM 0.9156 |
| Bakery | SSIM 0.9863 | SSIM 0.8854 |
SSIM measures image similarity, not human identity or a percentage of preserved content. Ask which reference image a quality claim uses: the original, or an intermediate that was already changed?
What if a detector misses the original too?
Then a low output score cannot establish that processing helped. In a separate Deprint pilot, UniversalFakeDetect missed 11 of 13 untreated AI-labelled scenes using its prescribed center crop. Another model, NPR, flagged two of four historical photographic controls.
Those small samples do not establish population-wide accuracy or false-positive rates. They show why every comparison needs suitable controls: known generated originals and known photographs, scored using the same procedure.
How to read a detector comparison
Look for four things: the exact files, the detector and date, the original-file results, and whether the detector influenced the edits. A test on new images with an independent evaluator is stronger evidence than a favorable score from a tool's own tuning loop.
Do not average unrelated percentages into one “AI probability.” Do not count a timeout or unavailable check as a pass. If a file changes after testing, its old result no longer describes the new bytes.
For a practical example involving different kinds of detection, read SynthID checks versus visual AI detectors.
Three common reasons a comparison can mislead
First, the tools may measure different targets. A watermark check asks about a supported embedded signal. A whole-image classifier asks about learned visual patterns. A face-manipulation model can focus on another task. Calling all three “AI detectors” does not make their outputs equivalent.
Second, the file paths may differ. One service may receive a full-size PNG while another receives a resized JPEG. Even if both files depict the same scene, that is not a clean same-file comparison. Use hashes and dimensions to track the inputs.
Third, the interpretation may differ. Two scores can fall on different sides of two vendors’ thresholds. A 0.6 from one model and a 0.6 from another are not necessarily comparable quantities.
| Before comparing results | Why it matters |
|---|---|
| Identify the task each model claims to perform | Similar names can hide different targets |
| Confirm the exact submitted file | Export differences can change the input |
| Record both score and label | A lower score can remain positive |
| Record tuning exposure | A model used to guide edits is not independent |
| Keep unavailable checks separate | Missing evidence is not a negative result |
Why a majority vote is not a universal answer
Two negatives and one positive are not automatically proof that the positive detector is wrong. The systems may share weaknesses, use overlapping training data or operate on different definitions. Without independently established origin labels and a suitable evaluation, a vote count cannot settle their accuracy.
For a known generated image, a negative visual result is a missed detection under that task’s definition. For a known camera photo, a positive result can be a false alarm. For an unknown file, the scores remain evidence to weigh alongside source history.
The local-target experiment is especially instructive because the two successful checks helped guide the edits. Counting those two as independent votes would exaggerate the strength of the result. The untouched Hive check supplied different evidence.
What “independent” should mean in a useful study
An evaluator should not influence candidate selection if it is meant to test transfer. Choose outputs using a documented rule before consulting that evaluator, and avoid repeatedly adjusting the method after every external score while continuing to call the test held out.
Hold out images too. A detector that was untouched can still be tested only on scenes selected after extensive exploration. Independent evaluation is stronger when both the relevant inputs and the evaluator were protected from that selection process.
These controls do not make a small study universal. They make its limits easier to understand. Report the selection procedure, the number of independent scenes, all completed checks and the visual quality outcome. That lets readers judge whether disagreement reflects a narrow optimization result or a broader pattern.
For help interpreting the percentages themselves, see what a 99% AI score means.
How we tested, and what remains unknown
The targeted-edit experiments ran October 7–8, 2026, on two reused Google-origin scenes. Eight exports in the follow-up were variants of those two scenes, not eight independent images. The files for external checking were selected by a fixed rule before viewing their Hive results.
The models, preprocessing and selection rules are documented in the public study methods. These observations establish a failure to transfer to Hive on those examples, not that every detector-targeted method always fails.