What is C2PA, and why AI images carry it

We opened 20 AI images from five models. All 20 carried a signed C2PA record. One plain re-save erased it.

Head of Content · Updated November 19, 2026 · How we test
The short answer

C2PA is an open standard from the Coalition for Content Provenance and Authenticity. It attaches a signed record, called Content Credentials, to an image, video or audio file, saying which tool made or edited it and whether generative AI was used. Instagram and TikTok read it to label AI content. It lives in the file's metadata, so a re-save or a screenshot can lose it.

Make your first photos

C2PA is an open technical standard that attaches a signed record, called Content Credentials, to a file. The record says which tool created or edited the image or video, including whether generative AI was used, and the signature lets anyone check that nobody changed the file or the record afterwards. Platforms read it to label AI content: it is one of the signals behind Instagram's AI info label, and TikTok uses it to auto-label uploads, as covered in TikTok's rules for AI-generated content.

Below: a test on 20 real AI images, what the record holds, how to check a file, and what PixelPig does with the record when you download.

Our test

On 8 November 2026 we sent one prompt, a night photo on a city bridge, to five image models through fal's API, four images each. The files went to our image storage byte for byte. On 17 November we opened all 20 with the open source c2pa-python reader (version 0.39.0) and printed who signed each record and what it claims.

ModelFiles with a C2PA recordSigned byTool named in the recordActions logged
Nano Banana Pro (Google)4 of 4Google LLCGoogle C2PA Core Generator Librarycreated as AI media, edited to apply a SynthID watermark
GPT Image 2 (OpenAI)4 of 4OpenAI OpCo, LLCOpenAI Media Service APIcreated as AI media, converted, watermarked
Seedream 4.5 (ByteDance)4 of 4Byteplus Pte. Ltd.c2pa-toolcreated as AI media
Wan 2.7 Pro (Alibaba)4 of 4fal - Features & Labels Inc.fal.aicreated as AI media
FLUX.1 dev (Black Forest Labs)4 of 4fal - Features & Labels Inc.fal.aicreated as AI media

"AI media" is the IPTC digital source type trainedAlgorithmicMedia, which IPTC defines as "Created using Generative AI". Every one of the 20 records carried it. None of the 20 files had camera EXIF data.

Three things stood out.

The signer is not always the model maker. Google, OpenAI and ByteDance's cloud arm signed their own outputs. Wan 2.7 Pro and FLUX.1 dev came back signed by fal, the service that ran them. The record names whoever ran the model, so the same model run elsewhere can carry a different signer or no record at all.

The signatures held. The reader marked all 20 records valid: the hashes matched the pixels and the signatures matched the records. It also flagged each signer as "untrusted", because we ran it without loading a trust list. That flag is a setting on our side, not a verdict on the signer.

The record is fragile. We opened one image from each model and saved its pixels as a new PNG, which is what many editors, a canvas redraw or a screenshot do. All five copies came back with no record at all. Then we flipped a single byte of image data in two originals and left the record alone. The reader kept the record but marked both files invalid, with a data hash mismatch. A record can be lost, but it cannot be moved to a changed file without the change showing.

What we did not test: how Instagram or TikTok label these files after upload. We did not post them.

What PixelPig does with Content Credentials

This part is described from our own code, not from a test upload.

Download. Every download first checks the file for a C2PA record (hasContentCredentials in src/utils/mediaProcessor.ts). A file that has one is saved byte for byte, under the extension of its real format, because a C2PA record signs the exact bytes and any change would break it. That rule came in on 16 November 2026. Before it, every image download was redrawn and lost its record. A file without a record is still cleaned: images are redrawn into a fresh PNG, and desktop MP4 downloads have their metadata removed. On phones, videos skip that step.

Generation. None of our server functions write a C2PA record of their own, and our download code notes that images from our own trained-persona pipeline carry none, so they download cleaned like any other unsigned file. Whether an output carries one depends on the model and on the service that ran it, as the test table shows. Our Terms of Service (section 6.4) say PixelPig may embed watermarks, metadata or C2PA Content Credentials, and that we do not promise a marking is applied to, or survives in, every output.

Your side. Our Acceptable Use Policy (section 7) says you may not remove, alter or obscure provenance marks, and that where the law asks you to disclose AI content, that duty is yours. The scheduler helps with the label: every Instagram and X post it sends asks for an AI-generated flag, and TikTok videos go out with TikTok's AI label set (the scheduling chapter of the Academy shows the TikTok side). TikTok photo posts you label in TikTok yourself. The full text is on our Terms and Acceptable Use pages.

How a C2PA record works

Picture a sealed envelope stapled to the photo. Inside are statements, which the spec calls assertions: "created by tool X", "converted", "AI media". The tool signs the envelope with a certificate. The envelope also holds a hash of the image itself, called the hard binding.

A verify tool checks two things. Does the signature match the record? Does the hash still match the file? Our byte flip failed the second check. An edit to the record would fail the first.

What the envelope cannot do is stay attached through every handling. Editors, messaging services and screenshots often write a new file without it, as our re-save showed. The spec fills that gap with soft bindings: an invisible watermark or fingerprint in the pixels that lets a service look the record up again. GPT Image 2's record logs a "watermarked" action and Nano Banana Pro's names a SynthID watermark, which points to that approach.

FactSource
The C2PA was announced on 22 February 2021 by Adobe, Arm, BBC, Intel, Microsoft and TruepicC2PA founding press release
Version 1.0 of the specification came out in January 2022C2PA announcement
The spec's version history lists 2.4 as April 2026C2PA Technical Specification 2.4
A manifest is a set of signed assertions about the fileC2PA Explainer
Every manifest must have a hard binding that ties it to the exact fileC2PA Implementation Guidance
Soft bindings, fingerprints or invisible watermarks, help find a manifest after the metadata is goneC2PA Soft Binding API
The "CR" icon shows a summary of origin, including whether generative AI was usedContent Credentials
Deployment Guidance 1.0 (8 July 2026) points to verify.contentauthenticity.org for checking filesC2PA Deployment Guidance
IPTC "trainedAlgorithmicMedia" reads "Created using Generative AI"IPTC NewsCodes, 2024 Q3
Meta reads C2PA and IPTC AI fields to label images (6 February 2024)Meta newsroom
TikTok uses Content Credentials to auto-label AI uploads (May 2024)TikTok newsroom

Method: checking a file for Content Credentials

  1. Keep the original download. Every edit or re-save is a chance to lose the record.
  2. Upload it to verify.contentauthenticity.org. It shows whether credentials are present, who signed them and whether they are valid.
  3. Read the result the way the C2PA says to. Valid credentials show the history. "No Content Credentials found" means the tool did not write any or they were lost, not that the image is real.
  4. Disclose anyway. Platforms want the label on realistic AI content whatever the file says.

If you make photoreal stills in the AI image generator, the craft side is in making AI photos look real. Better craft does not change the disclosure rule.

C2PA compared with other signals

SignalWhere it livesSurvives a re-save?Who reads it
C2PA recordSigned block in the fileNo, 0 of 5 in our testMeta, TikTok, verify tools
IPTC digital source typePhoto metadata, or inside the C2PA recordOnly if the editor copies metadataMeta
Invisible watermarkIn the pixelsDepends on the watermarkThe company that wrote it
Camera EXIFMetadata blockOnly if the editor copies itNot named by Meta or TikTok as an AI signal
Creator labelOn the platform postNot part of the fileEveryone who sees the post

Common mistakes

Thinking C2PA detects AI. It does not look at the picture. It records what a tool said about the file when the tool made it.

Treating a clean file as proof of anything. A missing record is the normal state for most files online, and one re-save is enough to get there.

Editing a signed file and expecting the record to follow. Most editors write a new file. If the record matters to you, post the original download.

Leaning on metadata to do the disclosing. The record gets lost; the platform label stays. For an account built around a persona, creating an AI influencer step by step puts disclosure in the setup from day one.

Assuming every route to a model writes the same record. In our test the same kind of record came from three model makers and one hosting service. Run the model somewhere else and the signer can change.

Bottom line

C2PA is a signed record inside a file that says how it was made. All 20 AI images in our test carried one, each marked as AI media, and one plain re-save erased it every time. PixelPig now keeps the record when you download a signed file. Post that original, and still add the platform's AI label yourself, because the label is what followers see.

Questions

What does C2PA stand for?
Coalition for Content Provenance and Authenticity. Adobe, Arm, BBC, Intel, Microsoft and Truepic announced it on 22 February 2021 as a Joint Development Foundation project.
What are Content Credentials?
Content Credentials is the public name for the C2PA record attached to a file. It holds signed statements about how the file was made and edited, and the "CR" icon marks files that carry it.
How do I check if an image has Content Credentials?
Upload the original file to verify.contentauthenticity.org, the tool the C2PA's deployment guidance points to. It reports whether credentials are present and whether they are valid. Developers can read the same record with the open source c2pa libraries, which is how we ran our test.
Does a missing C2PA record mean an image is real?
No. The C2PA's own guidance says a file without credentials either came from a tool that does not write them or lost them along the way. In our test, one re-save removed the record from every AI image, so a missing record proves nothing either way.
Does a screenshot remove C2PA data?
Usually. A screenshot is a new file built from pixels, and in our test a new file made from the same pixels carried no record. The C2PA describes watermarks and fingerprints, called soft bindings, as the way to find the record again.
Is C2PA the same as a watermark?
No. C2PA is a signed metadata record. A watermark is a pattern hidden in the pixels. The spec lets a record point to a watermark, and two of the five models we tested logged that they had added one. Nano Banana Pro names SynthID, and GPT Image 2 logs a watermark action.
Does PixelPig keep Content Credentials when I download an image?
Yes. Since 16 November 2026 a file that carries a C2PA record downloads byte for byte unchanged. Only files without a record are cleaned of other metadata. Our Acceptable Use Policy also says you may not remove provenance marks.
Do I still need to disclose AI content if my image has a C2PA record?
Yes. Meta and TikTok ask creators to disclose realistic AI content whatever the file carries, and where the law requires disclosure, PixelPig's Acceptable Use Policy puts that duty on the person who publishes. The record helps the platform; the label is still your job.

Make photos with your AI model

Make your first photos

Start here

Read next