Content Credentials and C2PA for Artists
Content Credentials are a tamper evident record, built on the open C2PA standard, that travels with an image and lists who signed it, which software touched it, what actions were taken and which earlier files it was built from. A verifier checks the signature and a hash of the pixels. They show history rather than truth.
If you have seen a small CR badge on an image, or a panel listing how a picture was made, you have met Content Credentials. They are built on an open standard from the Coalition for Content Provenance and Authenticity, shortened to C2PA, and the idea is direct: attach a signed record to a file that says where it came from and what was done to it. For an artist who wants to show a piece was made by hand, content credentials are a useful layer, provided you know exactly what the record contains and where it stops. This tutorial sits in the process and provenance pillar beside the process record it complements.
What are Content Credentials
A Content Credential is a small signed record, called a manifest, stored inside the image file or hosted alongside it. The manifest holds a claim, a set of assertions the claim points to, and a signature over the claim made with the signing certificate of the app that wrote it. One of the assertions is a hash of the image content, the hard binding. If anyone changes the pixels without writing a new manifest, the hash no longer matches and a verifier reports it.
The Content Authenticity Initiative FAQ describes the result as a nutrition label: it can say who produced the content, when, and with which tools and edits, but it is not designed to say whether an image is real. Identity is opt in, and including your name is always optional.
C2PA does not say whether an image is good, true or original. It says what the signer claims about its history, and whether that claim or the pixels have changed since.
What a C2PA manifest records
The C2PA 2.1 specification defines a set of standard assertions. The ones an artist will meet are below. The whole claim is signed once by the app that generated it, so the who signs it column is about which party stands behind each piece of data.
| Assertion | What it says | Who signs it | What strips it |
|---|---|---|---|
| Actions (c2pa.actions) | Created, opened, cropped, adjustedColor, drawing, placed; a standard manifest starts with created or opened | The app that wrote the manifest, with its certificate | Any re-encode or screenshot; cannot be redacted under the spec |
| Hard binding (c2pa.hash.data) | A hash of the exact pixels the claim covers | The same app | Any pixel edit breaks the match; cannot be redacted |
| Ingredients | Earlier files the image was built from, each with its own manifest if any | The app, recording what it opened or placed | Flattening to a new file without a manifest |
| Creator identity | A name or account the creator chose to attach, optional | The app, or an identity provider the creator used | Re-encoding; platforms may also choose not to display it |
| Digital source type | An IPTC value stating how the media was made | The app or generator that set it | Re-encoding; absent if the tool never wrote it |
| Time stamp | Proof the signature existed at a given moment | A trusted time stamp authority countersigning | Re-encoding |
Two details matter for artists. Redaction exists for privacy: a later manifest may remove sensitive assertions, but never the actions list or the hard binding, so a redacted credential still shows the editing history. And the standard defines soft bindings, a fingerprint or invisible watermark, so a verifier can try to re-find a manifest after a platform has removed it.
How the standard labels generated imagery
The standard carries a digital source type field that uses the IPTC Digital Source Type vocabulary. The value trainedAlgorithmicMedia is named Created using Generative AI. compositeWithTrainedAlgorithmicMedia covers a real image edited with generative tools, and humanEdits, added in September 2024, covers correction by a person with non generative tools. Note that digitalArt, the value earlier guides recommended for hand made digital painting, was retired on 2024-09-17; the vocabulary now says to use digitalCreation instead.
Several image generators attach credentials that declare their output as generated. A declaration exists only when the tool writes it and nothing downstream strips it, so the absence of a generated label is not evidence of anything.
How to read a credential with the public Verify tool
The Content Credentials Verify tool is a browser page that reads a file and shows its manifest. Use it on your own exports and on any image you are about to credit or question.
Load the largest copy you have
Drag the file onto the page. A screenshot or a re-saved copy shows no credential, so start from the original export or the platform download.
Read the signer first
The panel names who issued the credential. A known camera maker, creative app or studio account is meaningful; an unfamiliar signer tells you little.
Read the actions list
Look for the first action (created or opened) and what followed: drawing and adjustedColor suggest painting, placed means an ingredient was added, and a generated source type is a direct answer.
Check the match status
The tool reports whether the pixels still match the hard binding. A mismatch means the file changed after signing, which ordinary resizing can cause; ask about it rather than treating it as guilt.
Open the ingredients
Each ingredient may carry its own credential. Follow the chain back to the earliest signed file; that is the closest thing to an origin the record offers.
What happens to Content Credentials on upload
Most platforms re-encode images on upload, which produces a new file. Whether the manifest survives depends on whether the platform copies it across or hosts it remotely, and the CAI FAQ notes that even an included creator name may not be displayed, depending on the publishing or viewing platform. Some sites show a credential icon on images that carry one; many still discard embedded data entirely.
Because behavior differs by site and changes over time, test rather than assume. Upload a signed export, download it back with the platform's own save option, and drop the download into Verify. The same routine, extended to creator and rights fields, is the upload survival test in image metadata for artists.
What Content Credentials cannot prove
- Absence proves nothing. Most images online carry no credential, including most real paintings. A missing credential is not evidence of generation.
- A signature is a claim. It tells you who signed and that nothing changed since, never that the signer was honest; the CAI notes the name shown may differ from the owner of the work.
- Stripping is routine. A screenshot, a re-save or a platform re-encode removes the manifest without malice.
- It records software actions, not intent. A drawing action says a brush tool was used. It does not see the decisions behind the strokes; the archive in documenting your art process is what shows those.
- A broken match is not a verdict. Resizing for the web breaks the hard binding as surely as a forgery does.
How an artist can use Content Credentials well
Check what your tools support
Some creative apps attach Content Credentials on export, with an option to record your name and editing history. Look in the export or file info settings of the apps you already use.
Attach credentials to the final export only
Sign the file you publish. Keep the layered master and timelapse as the deeper record behind it, and do not sign work in progress you may later change.
Decide what identity to include
Identity is optional. A studio name or portfolio account is enough; the credential still works with no name at all.
Verify your own export, then the platform copy
Run the file through Verify before upload and after download from each site you use. Note which sites keep the manifest.
Pair it with a visible note
Because credentials get stripped, add a short plain language process note beside the work: tools, time taken, and that a timelapse is available on request.
Mistakes artists make with Content Credentials
| Mistake | Why it hurts | Fix |
|---|---|---|
| Treating a missing credential as proof of AI | Most hand made work has none; the claim collapses under the first counterexample | Say only what the record says; use the process archive for the rest |
| Signing a file, then resizing it for the web | The hard binding breaks and Verify reports a mismatch | Resize first, sign the final export last |
| Relying on the credential alone | One re-encode and the evidence is gone | Keep the layered master, session saves and replay as the durable record |
| Setting the retired digitalArt value | Modern readers expect digitalCreation | Use the current IPTC vocabulary value |
| Reading only the CR badge | The badge says a credential exists, nothing more | Open Verify and read signer, actions and match status |
Content Credentials as one layer of evidence
Content Credentials give the file you publish a signed, checkable history: signer, actions, ingredients and a hash that reveals later edits. Read them with Verify, sign only the final export, and expect uploads to strip them. The strongest position is a credential on the published file, a visible process note, and a private archive behind both. To use the same reading skills on an image that is not yours, continue with how to check where an image came from.
Common questions
What is a content credential?
A Content Credential is a signed manifest, defined by the C2PA standard, that travels with a file and lists who signed it, the software used, the actions taken (created, cropped, drawing, placed and so on), earlier ingredient files and a hash of the pixels. A verifier checks the signature and hash, so later edits to the data or the image are detectable.
How do I check content credentials on an image?
Open the public Verify tool at contentcredentials.org and load the largest original copy of the file. Read the signer, the actions list, the match status and any ingredients. A screenshot or re-saved copy will show nothing, so start from the original export or the platform's own download.
Do content credentials prove an image is not AI?
No. They show what the signer claims about the file's history and whether it changed since. A generated source type is a direct answer when present, but most images carry no credential at all, so absence proves nothing. Pair the credential with a process record and read the limits of AI image detectors before drawing conclusions.
How do I turn off content credentials?
Credentials are opt in, so they exist only when an app is set to attach them on export. Look in the export or file info settings of the app and disable the option, or export without your name while keeping the editing history. Identity is always optional under the standard.
Sources and further reading
Primary references for the facts and definitions in this piece, all from independent publishers:
- C2PA Technical Specification 2.1: manifests, signatures and hard bindings.
- IPTC Digital Source Type vocabulary: trainedAlgorithmicMedia, digitalCreation and the other source values.
- Content Credentials: Verify: a public tool for reading a file's credential.
- Content Authenticity Initiative: FAQ: opt in identity, tamper evidence and what credentials do not decide.
This method is part of the process and provenance pillar. Pair it with How to Document Your Art Process and How to Tell If Art Is AI: A Craft-Based Method.