Is this image AI? Here's what the file can actually tell you
This page does not give you a percentage. It reads the AI markers a file genuinely carries — signed content credentials, a source-type declaration, a generator's own name — and it tells you plainly when there is nothing there.
Run one image through three detectors
You will get three different answers and three different confidence percentages. That is not a bug in any one of them. It is what happens when you ask several systems to estimate a property that is only weakly present in the data they measure.
| Approach | What it measures | Why it fails |
|---|---|---|
| Frequency-domain detectors | Periodic artefacts left by upsampling inside the generator | Current generators have largely cleaned these up; the strong signals that remain come mostly from older models. |
| Classifier networks | A model trained to separate real from generated on a fixed dataset | It learns the generators in its training set. Newer models fall outside it and accuracy drops sharply. |
| Reading the file | A statement the file makes about itself | Does not fail — but only exists when someone wrote it. Absence proves nothing. |
If a detector is confident about a modern image, ask whether it is reading the file or guessing at the pixels. Reading the file is reliable. Guessing is not.
Three kinds of marker, in order of strength
1 · C2PA content credentials
A signed, tamper-evident record of where the file came from and what was done to it. In a
JPEG it lives in an APP11 segment as a JUMBF box; in a PNG it is a caBX chunk.
When present it is the strongest evidence available, because altering it breaks the
signature.
Full C2PA manifests are CBOR-encoded. This tool extracts the readable strings — it does not fully decode the manifest, and it says so rather than pretending.
2 · IPTC Digital Source Type
A single field that says what kind of content this is: photographed, digitally created, trained algorithmically, composited. When a generator or a newsroom sets it, it is a direct statement rather than an inference.
3 · Generator signatures in the text fields
Some tools write their own name into Software, XMP or the comment
field. If Midjourney or Stable Diffusion appears there, that is a fact you
can quote.
What a negative result means
Nothing was found is not the same as not AI-generated. Every major platform strips metadata on upload, and most image generators never write any. A clean result most often means the file passed through a social network.
To get a useful reading, use the file you received — the original download, not a screenshot and not a re-upload.
The one thing this page will not do is invent a number. A confidence score the tool cannot stand behind is worse than no answer at all.
Remove what you find
If the file is yours and you would rather not carry these markers, the same reader powers the remover on the main page — C2PA, XMP and IPTC all come out in one pass.