What Is C2PA? Content Credentials in AI Images, Explained
By AI Metadata Remover · Published 2026-10-21 · Updated 2026-10-21
C2PA is an open standard for attaching a signed record of a file's origin to the file itself. It is the technology behind “Content Credentials”, and tools such as ChatGPT and Gemini add it to the images they generate. This guide covers what the record contains, where it is stored, what each major generator really writes into a file, which platforms read it, what it cannot prove, and how to check your own images.
What is C2PA?
C2PA is the Coalition for Content Provenance and Authenticity, an industry group whose specification is maintained as a project of the Linux Foundation's Joint Development Foundation. The record it defines is called a C2PA Manifest, and “Content Credential” is the preferred everyday term. A related effort, the Content Authenticity Initiative, builds open-source tools around the standard.
It borrows a familiar idea: the same kind of digital-signature technology that secures HTTPS and PDF signing, applied to media files.
What is inside a manifest?
- Assertions: statements about the file, such as who or what created it, which tools edited it, and whether AI was involved.
- A signature: made with the private key of the software or hardware that performed the action, using standard X.509 certificates. Anyone can check it with the public key.
- A hard binding: a cryptographic hash of the content recorded at creation. If the pixels change, the hash no longer matches, so tampering is detectable. The specification calls this “tamper-evident”, not tamper-proof.
- An optional soft binding: a watermark or fingerprint that lets the manifest be looked up in cloud storage even if it was stripped from the file. A hard binding plus a soft binding is what the standard calls a Durable Content Credential.
A credential counts as valid when it is well-formed, unmodified and signed by someone on a recognised trust list. Valid does not mean true.
Where is it stored in the file?
The manifest is usually embedded in the image inside a JUMBF container. Where it sits depends on the format:
| Format | Location | What we saw |
|---|---|---|
| PNG | caBX chunk | ChatGPT PNG: one chunk of 23,654 bytes, about 1% of a 2.7 MB file |
| JPEG | APP11 segment | Nano Banana JPEG: one segment of 13,778 bytes |
| WebP | C2PA RIFF chunk | Our tool detects and removes this block |
Because it sits in a separate block, it can be read or removed without touching the image data. It can also be linked externally instead of embedded.
What each AI generator writes into the file
These findings combine public documentation with our own test files (two per generator, so treat the sizes as examples, not rules).
ChatGPT and DALL·E (OpenAI)
OpenAI began adding C2PA to DALL·E 3 images in early 2024 and says images from ChatGPT and its API carry it. Our PNGs held only that manifest: a signed OpenAI record plus the IPTC label trainedAlgorithmicMedia. No EXIF and no prompt. OpenAI itself says the metadata can be removed easily, by accident or on purpose, that most social platforms strip it, and that a screenshot can lose it. A missing tag therefore means “unknown”, not “not made by ChatGPT”. See our ChatGPT image metadata remover.
Gemini and Nano Banana (Google)
Google announced in November 2025 that Nano Banana Pro images made in the Gemini app, Vertex AI and Google Ads carry C2PA, with more products planned. Our JPEGs had one 13,778-byte block referencing Google, plus the same trainedAlgorithmicMedia label.
Google also uses SynthID, an imperceptible watermark in the pixels, introduced in 2023. Since late 2025 you can upload an image to Gemini and ask whether it was made with Google AI. SynthID is not C2PA: it lives in the image data, so removing metadata does not remove it. One hands-on test reported that a cropped Nano Banana image was not recognised, which suggests heavy edits can defeat detection, but that is a single report. See our Nano Banana page.
Midjourney
Public reporting and our files agree that Midjourney sets the IPTC DigitalSourceType to trainedAlgorithmicMedia. Our two JPEGs contained a small XMP block with just that label and a unique image ID, and no prompt and no C2PA. One third-party guide claims Midjourney writes the full prompt into a description field. We did not find that in our files, so check yours. See our Midjourney page.
Stable Diffusion
There is no single behaviour; it depends on the app that saved the file. AUTOMATIC1111 reportedly writes prompts and settings (steps, sampler, CFG scale, seed, size, model hash) into a PNG text chunk named parameters, and has a setting to turn this off. ComfyUI embeds its workflow as JSON in PNG text chunks, and that data can be missing if it was disabled or another program re-saved the file. Our two JPEG downloads from a web generator carried only a colour profile. PNG text chunks do not exist in JPEG, so some tools use EXIF fields instead. See our Stable Diffusion page.
What happens to it online?
- Many platforms strip it on upload, because they re-encode the image. Secondary sources list Instagram, X, Facebook and TikTok, and OpenAI says the same of most social platforms.
- Some platforms read it before stripping. Meta said in 2024 that it uses C2PA and IPTC signals to apply an “AI info” label on Facebook, Instagram and Threads, and noted that even light AI-assisted edits can trigger it. TikTok announced Content Credentials support in 2024, and LinkedIn displays credentials.
- Reading is not the same as keeping. Platform behaviour changes often, so test with your own file before relying on any of this.
What C2PA cannot do
The C2PA's own explainer is candid about its limits:
- It does not show that content is true, accurate or factual.
- Provenance can be incomplete, and adding credentials is optional. Absence should not be treated as untrustworthy.
- If a file is edited in a tool that does not understand Content Credentials, the edit may not be recorded.
- It is not DRM and does not restrict use.
- Metadata can be removed. Only a watermark or fingerprint can help recover it.
The rules that are coming
Article 50 of the EU AI Act requires providers of generative AI systems to mark synthetic images, audio, video and text in a machine-readable, detectable way. Commentary we reviewed says this applies from 2 August 2026 to new systems, with a grace period to 2 December 2026 for systems already on the market, and fines of up to €15 million or 3% of worldwide turnover. Sources disagree on some dates, so confirm them in the official text. The duty falls on the providers of AI systems, not on people who share an image.
Separately, platform rules and client contracts may require you to disclose AI use, whether or not metadata is present. This is general information, not legal advice.
How to check a file
- Web: the Content Credentials Verify tool accepts an upload or a public image URL and supports JPEG, PNG, WebP, AVIF, HEIC and more.
- Command line:
c2patool, open source from the Content Authenticity team, prints the manifest as JSON. - In your browser, privately: our free tool lists every metadata block it finds, with sizes, before you download a cleaned copy. Nothing is uploaded.
FAQ
Is C2PA the same as EXIF?
No. EXIF holds camera data and GPS. C2PA is a signed provenance record.
Is SynthID the same as C2PA?
No. SynthID is a watermark in the pixels. C2PA is metadata stored in a separate block of the file.
Does a missing manifest mean an image is real?
No. It means nothing either way. Platforms and apps strip metadata, and many tools never add it.
Does removing a C2PA block change image quality?
Not when the block is cut out without re-encoding the image. In our own test of an OpenAI PNG, the pixels were byte-identical afterwards.
Does removing it hide AI use from platforms?
Not reliably. Watermarks and detection models do not depend on metadata, and disclosure rules still apply.
Sources
See what is inside your own image
Our free tool lists every metadata block it finds before you download a clean copy. It runs in your browser, so nothing is uploaded.
Open the metadata remover →