What does the image say about itself?
Most AI image checkers guess from pixels. This one reads the file's own provenance metadata — Content Credentials, IPTC source type, generator fingerprints and camera EXIF — and tells you exactly what it found, or honestly says it found nothing.
What this actually checks
C2PA Content Credentials
The cryptographically signed provenance standard backed by Adobe, OpenAI, Microsoft, Leica and Sony. When present, it records how an image was made and edited — including whether a generative model was involved.
IPTC digital source type
A standard metadata field whose value 'trainedAlgorithmicMedia' is the publishing industry's explicit label for AI-generated imagery. Newsrooms and stock libraries rely on it.
Generator fingerprints
Stable Diffusion writes the entire text prompt into the PNG. Midjourney, DALL·E, Firefly, NovelAI and others leave their names in Software or CreatorTool fields. We surface all of it, including the prompt.
Read this before you trust any result
What metadata can prove
- +That a file declares itself AI-generated. A prompt embedded in a PNG or an IPTC 'trainedAlgorithmicMedia' tag is about as close to a confession as digital forensics gets.
- +That a signed C2PA chain exists, and what it records — including edits made after capture.
- +Which tool wrote the file, which is often more informative than a probability score.
What it cannot prove
- –That an image is real. Absence of metadata proves nothing — every major social platform strips it on upload, and so does a screenshot.
- –That the metadata is truthful. Metadata can be edited or copied from another file. Only a valid C2PA signature is tamper-evident; plain EXIF is not.
- –Anything about a screenshot, a re-encoded copy, or an image pulled from a social feed. In those cases you will nearly always get 'no provenance metadata'.
- –Anything from the pixels themselves. This tool does not run a visual model, so it will never flag an AI image that had its metadata stripped. That is a deliberate trade: we would rather return nothing than return a guess dressed up as a measurement.
Frequently asked questions
Why did a picture I know is AI-generated come back inconclusive?
Almost certainly because it was downloaded from social media, screenshotted, or re-saved. Any of those strips the metadata. The generator's fingerprint only survives in the original file as it came out of the tool. Ask whoever sent it for the original file.
Can this be fooled?
Plain metadata, easily — anyone can edit EXIF fields, and stripping them takes one command. A valid C2PA signature is much harder to forge because it is cryptographically signed, which is exactly why the standard exists. Read a positive AI result as reliable and a negative result as unknown.
Do you store the images I check?
No. We download the image, read its metadata in memory, return the result and discard it. We do not keep the file.
Why not just detect AI images from the pixels?
Because pixel-based image detectors are considerably less reliable than their marketing suggests, and they fail in the worst possible way — confidently. Metadata gives a verifiable answer or no answer. We would rather ship the verifiable half honestly than add another confident guess to the internet.
What about the text in a social post?
That is a different tool. Our social media AI detector reads a public post's caption and scores the writing. It also shows the post's image, but deliberately does not score it — this page is where image provenance lives.
Checking a whole social post?
Paste a post link and we will analyze the caption text too.
Try the social media AI detector