GENERATOR MARKERS · C2PA · SIGNALS — NOT VERDICTS

AI Image Provenance — follow the file's own trail

You cannot fingerprint pixels as "AI" with certainty — but files often leave provenance clues: generator strings, Content Credentials, digital-source declarations. PictureMatters surfaces those signals so you can weigh them with context. No detector drama; just evidence, labeled honestly.

The provenance signals PictureMatters looks for

  • Generator markers — recognized software strings from popular generators (OpenAI / DALL·E, Midjourney, Stable Diffusion, Adobe Firefly, FLUX and others) recovered from the file's bytes.
  • C2PA / Content Credentials — markers of signed provenance records. When present, they can carry creation and edit history with tamper-evident signatures.
  • IPTC digital-source-type — a formal declaration field, increasingly used to mark media as trained-algorithmic (AI) or composite.
  • Structure and entropy context — appended payloads, re-encode traces and entropy anomalies that add (or undermine) context.
  • Metadata survival — what survived the journey matters: a stripped file cannot testify about its origin at all.

What provenance can and cannot establish

Can: tell you what the file declares about its origin; show whether a signed credential exists; reveal platform or editor hops; flag files whose claims conflict (e.g., a generator string alongside camera EXIF).

Cannot: determine from pixels alone whether an image is AI-generated. There is no reliable pixel-level "AI detector," and claims to the contrary are marketing. Metadata — including AI markers — can be removed, and camera claims can be forged. The presence or absence of metadata is not, by itself, definitive proof of how an image was created.

Use provenance as one input next to source verification, reverse-image search and context. The AI provenance guide walks through a practical workflow, and What Is C2PA? explains the strongest signal available today.

Why "absence of markers" is the default

Screenshots, messaging apps, CMS re-encodes and quick exports routinely strip every trace. In practice most images you meet online carry no provenance at all — which is why PictureMatters reports metadata survival explicitly, and why a clean-looking file is best described as "no declared origin," never "verified human."

Frequently asked questions

Is PictureMatters an AI-image detector?

No. It reads provenance signals: generator strings, C2PA / Content Credentials markers, IPTC digital-source declarations, plus structure and entropy context. These are potential clues about a file's origin — not a verdict on whether pixels were machine-made.

If an image has no AI markers, is it human-made?

No. Most distribution paths strip metadata, and AI tools do not always embed markers. Absence of evidence is the normal state after sharing.

What are Content Credentials?

The consumer name for C2PA provenance — a cryptographically signed record of how a file was created and edited. When present and intact, it is much stronger evidence than ordinary metadata. See the C2PA Checker page.

Do AI images carry GPS or camera data?

Genuinely AI-generated images have no real capture event, so trustworthy ones do not claim camera GPS. If a file shows both a generator marker and a full camera EXIF set, that combination is itself a signal worth questioning.

Use the tool

Want to inspect an image? The PictureMatters analyzer reads metadata, structure, entropy and provenance signals in seconds — free, without an account.

Analyze an Image

Related pages