How accurate is AI image detection?
Short answer: good enough to be useful, not good enough to be authoritative. This page explains exactly where the line falls, so you can judge a result rather than just read it.
What this tool actually does
Four independent analyses run on every image, and their results are combined.
01 File forensics
This pass never looks at whether the picture looks like AI. It looks at how the file was made and encoded:
- EXIF, XMP and C2PA metadata — Camera make and model, exposure parameters, and any software tags. A complete capture block from a recognised camera leans toward authentic. A Software tag naming a generator, or a C2PA manifest declaring trainedAlgorithmicMedia, is close to conclusive.
- Container and dimensions — Compression structure, and dimensions landing exactly on multiples of 64 — typical of diffusion output grids.
- Pixel statistics — The absolute level of fine micro-texture, and a detrended autocorrelation test for the periodic ripple left behind by generative upsamplers.
02 Neural classification
A vision model trained on a large corpus of real and synthetic images examines the pixels and estimates the probability of synthetic origin. This is the stronger signal, and it determines the score whenever it is available.
03 How they are combined
Not by averaging, which would be a mistake. The classifier owns the score; forensics can only nudge it within a narrow band, and its ability to push a result down is deliberately larger than its ability to push it up — a full camera EXIF block is real information the model cannot see, whereas forensic findings pointing the same way as the model are usually the same artefact counted twice.
When the classifier is unavailable, forensics decides alone — and in that case the result is capped below the “likely” band no matter how strong the forensic findings look, and the result is labelled Advisory in the interface. That tier cannot justify a strong claim, so it is not permitted to make one.
Where accuracy holds, and where it collapses
| Image condition | Reliability |
|---|---|
| Unedited export straight from a generator | High |
| Original camera photo with EXIF intact | High |
| Downloaded from a generator, then resized once | Moderate |
| Passed through a social platform or messaging app | Low |
| A screenshot of an image | Low |
| Heavily retouched, denoised or filtered photo | Low — biased toward false positives |
| Stylised illustration, human or AI | Low |
| Human-painted over AI output | Very low |
The failure mode that matters most
A false positive on a real photograph of a real person does far more damage than a missed AI image, and the conditions that cause it are common rather than exotic:
- Screenshots. Screenshotting an image converts it to a PNG with no camera metadata and adds display-scaling artefacts that resemble generative smoothing. This is the single most common way a genuine photo gets scored as AI.
- Platform re-compression. Every social platform re-encodes uploads. That strips metadata, flattens the noise distribution, and leaves the file looking structurally like an export rather than a capture.
- Beauty filters and AI denoise. Modern phone cameras apply neural processing by default. The output genuinely was partly produced by a neural network, and it genuinely does look like generated imagery — because it partly is.
Practical rule: if the image came from a social platform, a screenshot, or a phone with computational photography enabled, treat any score as weak evidence. A high score in those conditions is not a finding.
What no detector can do
- Prove authorship. Detecting synthetic qualities is not the same as identifying who or what made an image, and it says nothing about whether a human directed, edited or repainted it.
- Keep up indefinitely. Detectors work because generators leave artefacts. Each generation of models leaves fewer, and detection is a moving target that is currently losing ground slowly. A score that is reliable today is not guaranteed to be reliable against next year's models.
- Justify a decision on its own. Not for a public accusation, not for moderation with real consequences, and not as legal evidence. Use it to decide where to look next.
Why there is no accuracy figure
You will see “99% accurate” on other detector sites. Treat any such number with suspicion, including from us if we ever quote one. A single figure requires a fixed test set, and the answer depends almost entirely on how that set was built: a corpus of clean generator exports produces a very impressive number that tells you nothing about how the tool behaves on a screenshot of a compressed JPEG — which is what most real queries are. Publishing one number would imply a precision that does not exist across that range, so this page documents the conditions instead.