PictureMatters
PictureMatters Editorial TeamPublished
GUIDE

How to Check for AI Image Provenance

There is no reliable pixel test for "made by AI." What you can do is read the file's own provenance signals — and know exactly what they can and cannot tell you.

The short answer

You cannot prove an image is AI-generated from its pixels. What you can find are provenance signals: generator strings embedded by some AI tools, C2PA Content Credentials from adopters, IPTC digital-source declarations, and metadata patterns that support or undercut a file's claimed story.

Signal 1: Generator markers

Several generators embed text markers — tool names, prompts, generation parameters — in the file. PictureMatters recognizes strings associated with major tools (OpenAI / DALL·E, Midjourney, Stable Diffusion, Adobe Firefly, FLUX and more). A marker is a genuine clue about the file's software origin — and it can be stripped like any metadata.

Signal 2: C2PA / Content Credentials

The strongest available signal. C2PA is the signed provenance standard adopted by Adobe, camera makers and several AI companies — credentials record creation and edits with tamper-evident signatures. If a file carries intact Content Credentials, they substantially constrain its history (assuming you trust the signer). Absence, again, proves nothing — most of the internet still strips it. See What Is C2PA?

Signal 3: IPTC digital-source-type

News-industry metadata increasingly labels media as "trained algorithmic media" (AI), a "composite", or a "digital capture". When present, this is a formal declaration — valuable because it is a claim by the publisher, distinct from anonymous strings.

Signal 4: The claim-conflict check

Some combinations are self-undermining: a "camera original" shot with rich EXIF and a generator string; GPS coordinates on an image whose tool marker says it was generated last week. The PictureMatters report lays all findings side by side so conflicts are visible.

Signal 5: Metadata survival

Context first: where did the file come from? Straight from a generator's download button, it may carry rich provenance. After WhatsApp or Instagram, assume everything was stripped — absence of markers there is meaningless. The report's metadata-survival meter grades exactly this.

What doesn't work

  • Vibes — "it looks too smooth" was weak in 2023; it is useless now.
  • Online "AI detectors" — published evaluations show high error rates in both directions; treat percentage claims skeptically.
  • Single-signal certainty — a marker without context, or absence without journey knowledge, is not a verdict.

A practical workflow

  1. Ask for (or find) the highest-quality original — provenance rarely survives re-sharing.
  2. Run it through the AI provenance checker and read generator, C2PA and survival findings together.
  3. Reverse-image search to find first publication.
  4. Check the poster's source and history.
  5. Conclude with calibrated language: "carries markers consistent with…" — not "is AI."

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

Frequently asked questions

Do AI images have EXIF?

Usually not camera EXIF (there was no camera). Some generators embed their own metadata or C2PA credentials; many strips or exports strip everything.

Can metadata prove an image is NOT AI?

Camera EXIF supports a "photograph" story but doesn't prove it — EXIF can be forged by injecting a generated image into a camera-file container. Convergence of evidence beats any single field.

Related guides

›How to Check If an Image Has Been EditedSoftware traces, structure and compression — what honestly helps and what does not.›What Is C2PA and Content Credentials?The open standard for cryptographically signed media provenance, explained.›Does WhatsApp Remove EXIF Metadata?What actually happens to photo metadata when you send it through WhatsApp.