guide · updated October 2026

What are C2PA “Content Credentials”?

If you've read about the “Made with AI” label, you've bumped into the term C2PA — sometimes dressed up as “Content Credentials” or “provenance data”. Here's the whole story in plain words: what it is, where it physically lives inside your files, who writes it, and whether it can be removed.

The one-paragraph version

C2PA (Coalition for Content Provenance and Authenticity) is an industry standard — backed by Adobe, Microsoft, Google and others — for stamping a file with information about its origin: what tool created or edited it, when, and whether AI was involved. Those stamps are called Content Credentials. Think of it as a paper trail glued to the file itself, meant to travel with the image wherever it goes.

Where it physically lives in your file

Content Credentials aren't some mystical cloud record — they're bytes inside the image file you have on disk:

• JPEG — stored in special markers: an APP11 segment holding JUMBF data, and/or XMP written into APP1. Related tags (IPTC “Digital Source Type”, plain EXIF) sit in APP1/APP13.
• PNG — custom chunks named caBX (C2PA box) and tEXt/iTXt/eXIf chunks.
• WebP — EXIF/XMP chunks inside the RIFF container, plus the “has EXIF/XMP” flags in the VP8X header.
• All formats may also carry loose provenance hints in trailing data after the image itself.

What the stamps actually say

The most consequential field for Instagram labeling is the Digital Source Type. Some values you may see:

• trainedAlgorithmicMedia — generated by an AI model (the big one)
• compositeSynthetic / compositeTrainedAlgorithmicMedia — stitched or AI-composited
• digitalCapture — a camera/phone shot (the good one)

So a file can literally carry the sentence “part of me was made by a trained model” — in machine-readable form. Instagram, TikTok and other platforms read exactly these fields when deciding to apply an AI label.

Who writes it into your files

• Generators: Adobe Firefly, DALL·E / ChatGPT images, Midjourney, Imagen — almost all major ones now stamp output by default.
• Editors with AI features: Photoshop (Generative Fill, and Content Credentials embedding on export), Lightroom, Canva, phone “magic eraser”-style cleanups.
• Server-side processors: some enhancers, upscalers and background removers append their own credentials to whatever they return.

Can Content Credentials be removed?

Yes — for the metadata itself. They're ordinary data sections in the file, and tools like Unflag delete them: the C2PA boxes, the XMP/IPTC/EXIF records, and trailing data, while leaving the compressed image pixels untouched (lossless). A fresh re-encode (the “shrink” tool, or Recompress mode) removes them as a side effect too.

Two honest limits:

• Pixel watermarks are different. Google's SynthID and similar techniques bake the signal into the pixels themselves; only recompressing the image meaningfully degrades those, and even then it's not guaranteed gone.
• Validation stops working, not the platform. If a claim needs to be proven later (press, stock, legal), a stripped file loses its receipt. Some photographers deliberately keep credentials; stripping is a choice, not an obligation.

Quick self-check for your files

Curious what your own images carry? Drop one into contentcredentials.org/verify — if it shows credentials, that's what platforms read too. Then, if you'd rather they didn't: wipe the file and re-check; it'll come back clean.