scriptiii
100% Client-Side

C2PA · IPTC · prompt blocks

AI metadata checker

See whether an image declares or signs itself as AI-generated — Content Credentials, IPTC source type, a stored prompt — then remove every marker and scan the result again.

The scan runs in this tab. Your image is never uploaded, and the cleaned copy is built from the pixels already in memory.

Short answer: drop the image. The report lists a C2PA manifest, an IPTC “trained algorithmic media” declaration, a stored prompt block or tool-name Exif — each with what the file literally says. Then download the cleaned copy, which is re-scanned and should report zero markers.

The three things people mean by “AI watermark”

Declared provenance is metadata a tool adds on purpose: C2PA Content Credentials, or an IPTC source type of trainedAlgorithmicMedia. Platforms read it, and it is removable because it is metadata. Tool fingerprints are accidental: a Stable Diffusion PNG stores the whole prompt, seed and sampler, and that block is a confession in plain text. Both of those this page finds and removes.

Pixel watermarks are different. SynthID and similar schemes encode a signal in the image data itself; they survive metadata scrubbing, and a tool that promises to remove them is either re-encoding and guessing or lying. This page does neither — it tells you which of the three you are looking at.

Screenshots, and why “no marker” is not “not AI”

A screenshot of an AI image has no manifest, no prompt block and no Exif. The absence of markers means the markers are gone, not that a person drew the picture. That distinction is the whole reason this page reports facts about the file instead of a verdict about the image.

AI metadata questions

What is actually checked?

Four families: a C2PA Content Credentials manifest (a signed box that can name the tool that made the image), an IPTC digital source type such as trainedAlgorithmicMedia in the XMP packet, a Stable Diffusion or ComfyUI text chunk that stores the prompt and seed, and ordinary Exif / IPTC blocks that name the software. Every hit is listed with what the file literally says.

Can it tell me whether an image is AI-generated when there is no metadata?

No, and no honest metadata tool can. Removing metadata is one keystroke, screenshotting is two, and a phone photo and an AI render look identical to a byte scanner. What this page can do is tell you whether the file still carries a declaration — that is a fact about the file, not a guess about the picture.

What about SynthID and other invisible watermarks?

They live in the pixels, not the metadata, so a metadata cleaner cannot remove them. The page says this next to the result instead of quietly implying the image is now untraceable. A re-encode can damage some of them and cannot be relied on.

Why does it re-encode instead of editing the metadata blocks?

Because editing blocks is how a marker survives: delete the Exif segment and the XMP copy still names the tool, keep one thumbnail and the GPS coordinates come back with it. The cleaner redraws the pixels into a new PNG, so every metadata segment is gone by construction — and then it scans the result again and shows you the count, which should read zero.

Related tools

EXIF removerGPS, camera and timestamp in photos.Hidden text charactersThe same idea for text files.PDF metadataProducer and XMP that name the tool.