Read the making-of history baked into the file for definitive evidence — not just a pixel guess.
Free with no sign-up, and the analysis runs inside your browser — files never leave your device. On our public gate (30 real + 83 AI images) it measured 94.7% at original quality and 87.6% on recompressed copies, with zero real photos confidently mislabelled as AI. Treat it as an indicator, not proof: the newest commercial generators evade every detector.
Read the "provenance" information left inside an image file to find clues about which camera, app or AI created it. It looks at C2PA (Content Credentials) together with the metadata.
C2PA (Content Credentials) is an open standard where cameras, editing programs and generative AI sign into the file "when, with what, and how it was made". Because it is a digital signature it is hard to forge, and when present it lets you verify the creation history with high confidence.
Some generative AI (DALL·E, Firefly) and the latest cameras and editors support this. When a signature explicitly states "AI-generated", it is strong evidence. But many tools still do not add it, so the absence of a signature does not mean "a human photo".
Even without a C2PA signature, EXIF and XMP metadata can reveal the camera model, editing-software name (e.g. Photoshop) or traces of a generation tool. A file with "no camera info at all, only editor traces" is a clue to suspect compositing or heavy editing.
Metadata, however, is easy to strip or alter, so it is not absolute proof. Combining it with the home page's AI detector, which analyses the pixels themselves, makes the judgment far sturdier.
Several places inside a file offer clues about how it was made. Beyond a signed provenance record, EXIF has a software field naming the editing program, and the XMP block sometimes carries the generating tool's own name. Some generation services write the prompt or model name straight in.
Invisible watermarking is spreading too — faint patterns embedded in the pixels, readable only with a purpose-built detector. Those marks are often verifiable only through the detector of the company that made them, and they do not survive heavy editing or re-compression.
This tool reads the values actually present in the file and shows them as they are. It does not infer, and it does not manufacture values that are missing. The result screen is a list of what was there, not a verdict about what the image is.
Here is the common misreading. Empty provenance feels suspicious, when in fact most ordinary photographs are empty. Anything that passed through social media, arrived by messenger, or is a screenshot has almost certainly had its metadata stripped.
The reverse holds too. Provenance present and a camera model listed does not guarantee a genuine photograph. Metadata can be edited in, particularly any field that carries no signature. That is why a signed record and a merely written value have to be read differently.
So use this check in pairs rather than alone. Where a provenance record exists, treat it as strong; where it does not, do not treat the absence as grounds, and move the question to the pixel check and reverse image search. The three are different kinds of evidence, and confidence comes from their overlap.
Files retaining provenance are uncommon, but when they do the record beats any other check available here: a signature is hard to forge and it names the tools involved. If you get a result, store it with the original file rather than as a screenshot.
Three settings account for most real use: a first pass on material received as a tip or for reporting, checking the origin of listing photos in a private sale, and establishing an editing history in a rights dispute. In all three, hand over the original file alongside this result rather than the result alone.
A file downloaded straight from an image generator carried C2PA credentials naming the tool that made it, signed. This is confirmation rather than estimation. Whatever the pixel score says, the file itself has stated that it is generated.
A DSLR original returned manufacturer, model, lens, aperture and capture time: a combination generators rarely produce. That is strong evidence for a real photograph, but not proof. Metadata can be written or edited by hand, so unlike signed C2PA it remains forgeable.
Screenshotting, resizing or a single upload-and-download round trip wipes C2PA and EXIF completely. Most files people bring here are already in that state. "No creation data" therefore proves nothing at all. When this panel is empty, move the question to pixel analysis and reverse image search.
No. Most generation tools still do not embed C2PA. When there is no signature, read it together with the pixel-analysis result.
It is easy to erase or tamper with, so it is only a clue. A C2PA digital signature is far more reliable.
It will run, but will usually be empty. Messengers re-compress images in transit and commonly clear metadata. Ask the sender for the original file — sent with an original-quality option or as a file attachment, the data may survive.
If it comes from a signed record, it is strong. Fields written without a signature can be added or removed by editing, so read carefully whether a displayed value came from the signed provenance record or from ordinary metadata.