What a Content Credentials manifest actually records
C2PA — the Coalition for Content Provenance and Authenticity — is an open technical standard for attaching provenance data to a file. The data structure is called a manifest: a compact record that travels with the asset, usually as a JUMBF box (a container format for metadata).
A manifest is not a single "yes, this is real" stamp. It is a set of claims, each signed by a key. A photo from a supported phone might carry a claim from the camera, a claim from the editing app, and assertions about what changed. A video might list trim, crop or redaction actions.
| Part of the manifest | What it can contain | Why it matters |
|---|---|---|
| Claim generator / signer | Camera, phone, editor, platform or publisher — plus the certificate used to sign | Tells you who is making the claim, not whether they are honest |
| Assertions | Created, edited, resized, cropped, redacted, and other actions | Gives an edit history, if the tool wrote one |
| Ingredients | Hashes and thumbnails of source assets | Links an output to the files it came from |
| Hashes | A cryptographic binding to the asset bytes | Lets a verifier detect changes made after signing |
| Timestamps | When the claim was created or signed | Useful context; not proof of when a scene happened |
Rule of thumb: a manifest records claims, not truth. It can prove that a file has not changed since a particular key signed it; it cannot prove the scene, the person or the event was real.
Where the manifest lives in each file type
The manifest is metadata, so it sits next to the pixels rather than inside them. That is why re-encoding, screenshotting or copying an image through a platform that strips metadata can make credentials disappear without touching the visible content.
| File type | Typical carrier | Notes |
|---|---|---|
| JPEG | APP11 segment (JUMBF) | The most common carrier for phone and camera photos |
| PNG / WebP | A dedicated chunk in the image container | Chunk name and layout are format-specific |
| MP4 / MOV | A box inside the ISO base media file format | Video manifests can include trim and edit actions |
| Plain text / OGG / AVI | Added in C2PA 2.3 (February 2026) | Alongside EXIF-preserved originals; the scope keeps growing |
| PDF / DOCX | Not part of the C2PA 2.3 list | Do not assume a document has credentials because of its extension; verify the specific file |
Detected vs verified — the distinction that matters
Most badges and viewers stop at the first step: detected. The bytes contain something that looks like a C2PA manifest. That is useful — it tells you there is provenance data to inspect — but it is not a judgment about whether the credential holds up.
Verified is the second step. A verifier parses the manifest, checks the signature chain, validates the certificate against a trust list, and compares the signed hash with the asset you actually have. A manifest can fail this for ordinary reasons: the file was edited after signing, the certificate expired, the signer is unknown, or the tool does not support the format.
Docfy's position: the metadata viewer flags that a C2PA manifest appears to be present and shows the standard metadata around it. It does not verify signatures yet. For verification, use an official verifier; we link to one below.
Three ways to check a file
Different tools answer different questions. Use the one that matches the decision you are about to make.
| Tool | What it checks | Where your file goes | Best for | Limit |
|---|---|---|---|---|
| Content Credentials Verify (official) | Manifest structure, signature chain, trust list and asset hash | Uploaded to the verification service | "Is this credential valid, and who signed it?" | Needs network; not a local read |
| Google SynthID Detector | Google's invisible SynthID watermark | Uploaded to Google; sign-in required | "Was this made by a supported AI tool?" | Not C2PA; limited to supported ecosystems and formats |
| Docfy metadata viewer | EXIF, XMP, QuickTime tags and whether a C2PA manifest appears present | Stays in your browser | "What does this file carry before I share it?" | Presence only; no signature verification |
The official verifier is the right tool when the answer has consequences. A local viewer is the right tool when you want to look without sending the file somewhere else.
C2PA vs Apple Reference Image vs SynthID
Three different mechanisms are often described as "provenance". They are not interchangeable.
| Mechanism | What it is | What it can show | What it cannot |
|---|---|---|---|
| C2PA / Content Credentials | Open manifest standard signed by creators, tools, devices or platforms | A signed chain of claims about origin and edits | Truth; and the manifest can be stripped or edited |
| Apple Reference Image (September 2026) | Sensor-level capture signature from a supported iPhone camera, plus Private Cloud Compute | That a capture came from that device's sensor path | Ecosystem-limited; not a general C2PA replacement |
| Google SynthID | Invisible watermark embedded by supported generative tools | That content was likely produced by a supported model | Not a signed manifest; requires Google's detector |
| Platform AI labels | Policy labels shown by YouTube, TikTok or Instagram | That the platform decided to label the item | Not technical proof; rules vary by platform |
Sources: Apple Security Research, Apple Reference Image (Sep 2026); Google Security Blog, Pixel 10 C2PA (Sep 2025); YouTube, AI labels (May 2026).
Why this matters in 2026
Provenance stopped being a nice-to-have when labeling became a legal obligation for some systems. The dates below are the ones that show up in search and in contracts.
| When | What changes | Who it affects |
|---|---|---|
| 2 Aug 2026 | EU AI Act Article 50 transparency duties apply; generative systems must mark outputs in a machine-readable way | Providers of generative systems in the EU |
| 2 Dec 2026 | Grace period ends for generative systems already on the market before 2 Aug 2026 | The same providers, for existing systems |
| 2 Aug 2026 | California CAITA's first duties take effect; SB 1000 (30 Sep 2026) explicitly allows qualifying third-party verification tools | Covered generative AI providers; platforms from 1 Jan 2027; capture devices from 1 Jan 2028 |
| May 2026 | YouTube moves AI labels into the player and automatically detects undisclosed realistic AI content; C2PA "fully generated" labels are permanent | Creators and viewers on YouTube |
Plain-English summary, not legal advice. The obligations sit with providers, platforms and device makers — not with every person who edits a photo. If you need to know whether your product is in scope, ask a lawyer, not a metadata viewer.
Sources: EU AI Act Article 50; Morrison Foerster on California SB 1000 (Oct 2026); YouTube AI labels (May 2026).
What to do when a file lands in front of you
Look before you trust
Open the file in a local viewer to see what metadata and provenance markers it carries. Presence is a pointer, not a verdict.
Verify when it matters
If the decision is publication, payment, evidence or compliance, run the file through the official verifier and read which key signed it.
Ask for the original
A screenshot, a re-encoded copy or a platform download often loses the manifest. The original file is the only one that can carry the full credential.
Cleaning is a deliberate choice
If you remove metadata to protect your own privacy, use the photo metadata remover or the PDF remover. Remember that stripping provenance also removes the evidence that could support you later; do not do it to pass off someone else's work.
What this guide and our viewer cannot tell you
- No signature verification. The viewer flags presence only; a green badge is not a certificate.
- Absence is not proof of fakery. Most files in the wild have no manifest at all, including genuine photos.
- Presence is not proof of truth. A valid manifest can describe a staged scene, an AI-generated asset, or a deliberately misleading claim.
- Metadata can be stripped or re-added. Treat any single file as evidence of what that file says, not of what happened.
- Format coverage changes. C2PA's scope keeps growing; our viewer supports PDF, JPG, PNG, WebP, MP4 and MOV today.
- Not legal advice. The dates and duties above are a summary; obligations depend on your role, market and product.
Questions people ask
Does C2PA prove a photo is real?
No. A manifest records signed claims about who made a file and how it was edited. A valid signature can prove that a claim was made by a particular key and that the file has not changed since signing — not that the scene, the person or the event was real.
Can C2PA metadata be removed?
Yes. A manifest is data stored next to the pixels, so re-encoding, screenshotting or stripping metadata can remove it. Removing provenance from your own file is sometimes legitimate, but doing it to deceive can violate platform rules or sector regulations. This page is not legal advice.
Do PDFs or DOCX files carry C2PA?
Do not assume they do. C2PA 2.3 (February 2026) added plain-text documents, OGG Vorbis, AVI and EXIF-preserved originals, but PDF and DOCX are not part of that 2.3 list. Check a specific file with a verifier rather than guessing from its extension.
What is the difference between detected and verified?
Detected means the bytes appear to contain a manifest. Verified means a verifier validated the manifest structure, the signature chain, the trust list and the hash of the asset. A file can contain a manifest that fails verification, for example if it was edited after signing.
Does Docfy verify Content Credentials signatures?
Not yet. The metadata viewer flags that a C2PA manifest appears to be present and shows the standard metadata around it. For signature verification, use the official Content Credentials Verify tool.
What about SynthID and platform AI labels?
SynthID is Google's invisible watermark for AI-generated content and is checked with Google's SynthID Detector. Platform labels on YouTube, TikTok and Instagram are policy labels, not a signed technical manifest. They answer different questions from C2PA.
Why does a file show a manifest but the verifier says it is invalid?
Common reasons are that the file was edited after signing, the signer certificate is not trusted or has expired, the manifest is truncated, or the verifier does not support the format. A failed verification does not prove the content is fake — it only means this credential cannot be trusted as-is.