Durable Content Credentials: how a watermark and a fingerprint recover a stripped record
Durable Content Credentials are an approach described by the Content Authenticity Initiative that combines three things: signed metadata, an invisible watermark and a fingerprint. If the signed record is stripped from a file, the watermark can point to a stored copy and the fingerprint can confirm the match. Our checker does not perform that recovery.
What are durable Content Credentials?
A normal Content Credentials record is signed metadata inside the file. The Content Authenticity Initiative's post on durable Content Credentials says that metadata of any kind can be removed on purpose or by accident, even though tampering with it can be detected. Its answer is to add two more layers next to the signed record.
The post says the C2PA specification has long included "soft bindings" for this kind of durability. As of April 2024 it described technology providers as just beginning to implement the durability approach. Our page on Content Credentials explains the basic record.
What are the three techniques?
| Technique | What it is | Limit the post names |
|---|---|---|
| Secure metadata | A signed record inside the file about how the content was made or edited. | It can be removed, on purpose or by accident. |
| Watermarking | Hidden data in the content that a watermark detector decodes. The post says state-of-the-art watermarks can withstand cropping or rotating images and noise added to video and audio, and can survive rebroadcasting such as screenshots. | It carries only a small amount of data, and it can still be removed or spoofed. |
| Fingerprinting | A short numerical code computed from the content (pixels, frames or audio) and stored separately. It is used later to find a match. Nothing is embedded in the media. | Matching is inexact, so it cannot be perfectly certain. |
How is a durable credential created?
The post describes five steps.
- A watermark is applied that encodes a short identifier and a pointer to where the record can be fetched.
- A fingerprint is computed from the media.
- The watermark identifier and the fingerprint are added to the Content Credentials record.
- The whole package is signed.
- The record is embedded in the file and also stored remotely.
The post says the pointer can lead to a Content Credentials cloud host or a distributed ledger. It adds that this approach is not right for every case: a photojournalist focused on privacy, for example, may not want anything stored on a server or blockchain.
How does recovery work when the metadata is missing?
The post lists four steps.
- Decode the watermark to get the identifier.
- Use the identifier to look up the stored record.
- Use the fingerprint to confirm the image is a perceptual match, which guards against a spoofed or misread watermark.
- Check the signature and the data in the recovered record.
What does Adobe say about recovering Firefly credentials?
Adobe's Content Credentials overview for Firefly (last updated August 25, 2026) says the credentials are attached to the file and "may be stored" in Adobe's public Content Credentials cloud. It says a stored copy makes them durable and recoverable with the Inspect tool.
The word "may" matters: Adobe does not say every file has a stored copy. Adobe's Inspect page invites you to drag in a file or a screenshot.
The pages we read do not say which watermark or fingerprint Adobe uses, or how often a recovery succeeds. Inspect is another company's site, not ours, and we do not control what it does with a file you give it.
Is TrustMark the watermark in durable Content Credentials?
The Content Authenticity Initiative's post does not name TrustMark. The TrustMark repository describes the project as an open source watermarking implementation for the Content Authenticity Initiative, and credits Tu Bui, Shruti Agarwal and John Collomosse as the authors of its paper. Its README includes a C2PA example that shows how to indicate a TrustMark watermark in a manifest, and it says the JavaScript version only decodes.
No page we read says Adobe's Firefly recovery runs on it, so we do not say so. Our checker does not run TrustMark.
What did we measure?
Our checker never fetches an online record and reads no watermark that points to one. We tested what that means.
| Test | Files | Result |
|---|---|---|
| A file whose metadata points to an online copy of its record | 5 files we made: JPEG, PNG, WebP, AVIF, GIF | All 5 read "Content Credentials are not available locally". For the JPEG, a test server on the pointed-to address received 0 requests. |
| Real files with a record, redrawn on a browser canvas and saved again as PNG and as JPEG (quality 0.92) | 9 real files, 18 outputs | 0 of 9 kept a record after the PNG, 0 of 9 after the JPEG. All 18 outputs read "No supported provenance signals found". |
The nine real files are 3 from Adobe Firefly, 2 from Microsoft Image Creator, 3 ChatGPT files and one C2PA test image. Redrawing on a canvas stands in for what a screenshot or a re-encode does to the pixels. A canvas keeps pixels only, so no record can travel with it.
The 18 outputs show that our checker found no record and no metadata in them. They do not show whether a watermark or fingerprint would still lead Adobe's tool to a stored record, because we do not read either, and we did not run Adobe's decoder on our files.
Limits: one browser, one JPEG quality, no resizing, and a canvas is not a platform. Our accuracy page explains what the checker cannot see, including a stripped record.
What this means when you check an image
Open the checker, choose the file, and look for one of two headlines.
"No supported provenance signals found" means we found no signed record, no source label in the metadata and no Stable Diffusion watermark pattern. The line under it says: "This does not mean the image is real or human-made." A stripped record is not recovered.
"Content Credentials are not available locally" means the file points to a record stored online. The line under it says: "We never fetch it, so nothing about your image leaves your browser. Adobe Inspect can look it up." It tells you a pointer exists, not what the stored record says.
If you want a recovery attempt, Adobe Inspect is the tool Adobe describes. Adobe says Firefly credentials "may be" stored, so an empty answer does not rule out that a record once existed.
Guides on this topic
- What are Content Credentials? is the main page on this topic.
- Why Content Credentials disappear covers the everyday steps that strip a record.
- Watermark, metadata or fingerprint compares the three ways to tie a claim to an image.
- Invisible AI watermarks compared lists which marks exist and who can read them.
- Adobe Firefly's Content Credentials explains the legacy trust-list result.
Frequently asked questions
- Do durable Content Credentials survive a screenshot?
- The Content Authenticity Initiative says the strength of a watermark is that it can survive rebroadcasting such as screenshots, which effectively remove signed metadata. It also says watermarks can be removed or spoofed. We could not test this: our checker does not read these watermarks.
- Does every Adobe Firefly file have a stored record?
- Adobe's page says Firefly credentials "may be stored" in its Content Credentials cloud. It does not say every file has one.
- Can your checker recover a stripped record?
- No. We do not decode a watermark that points to a record, we do not compute or match fingerprints, and we never fetch an online record.
- Is a recovered record proof that the image is authentic?
- No. A recovered record is still the claim of whoever signed it. Separately, the Content Authenticity Initiative's post says Content Credentials are not a silver bullet.