How Brands Use C2PA Content Credentials to Disclose AI-Generated Media
A brand publishes an AI-generated campaign image that carried a valid C2PA credential from its generation tool. Then a resize step re-encodes the file, and the LinkedIn post goes out with no AI badge.
A brand publishes an AI-generated campaign image that carried a valid C2PA credential from its generation tool. Then a resize step re-encodes the file, and the LinkedIn post goes out with no AI badge. To prevent that, brands treat C2PA as one layer of disclosure: preserve the signed manifest from generation, extend it at edit, and render a label from structured data at publish. In Sanity, AI-use fields in the Schema hold the disclosure, and Functions check and stamp each asset as it publishes.
The stakes are rising. EU AI Act Article 50(2) requires providers of generative AI systems to mark outputs in a machine-readable format from 2 August 2026. LinkedIn, TikTok, and YouTube already read Content Credentials to decide which posts get an AI label. When an AI-generated asset shows up with no disclosure, people don't read it as a technical glitch. They read it as concealment, and the brand pays for that in trust.
This guide treats a Content Credential as a fact your content operation has to keep true at every step, not a property of the file that you attach once. The file credential is fragile. What lasts is a disclosure that lives as structured data in your system of record, where it can re-sign, re-label, and audit every channel the asset reaches.
What does a C2PA Content Credential record about an asset?
A C2PA Content Credential records who made an asset, how it was changed, and whether AI was involved, all in a signed structure attached to the file. "Content Credential" is the plain-language name for a C2PA Manifest. The C2PA Explainer 2.2 describes a manifest as a cryptographically bound structure that records an asset's provenance. The manifest is made of assertions, and three kinds matter most for disclosure: origin (when and where the asset was created), modifications (what was done and with which tools), and use of AI (how the asset was authored).
AI generation has a specific machine-readable form. The manifest carries a c2pa.actions assertion. Inside it, a c2pa.created action sets digitalSourceType to the IPTC term trainedAlgorithmicMedia, which means media made entirely by a generative model. The specification also defines C2PA's own trainedAlgorithmicData value. LinkedIn and TikTok read this field to decide whether to apply an AI label, so for a brand that wants its disclosure to travel, it is the most important line in the manifest.
The manifest is signed. If someone changes the pixels or the assertions after signing, a verifier can detect the mismatch. A verifier also checks whether the signer is on a trust list, so a credential from a recognized signer tells the audience who vouched for the record.
What a Content Credential does not do matters just as much. It makes no judgment about whether the content is true. It only shows that the provenance data is well-formed, untampered, and signed. For brands, a Content Credential is a disclosure mechanism, not a fact-check. People still decide whether an AI-generated image is right for a campaign. The credential just records that a generative model was involved.
In practice, many AI assets reach a brand with a manifest already attached. OpenAI attaches Content Credentials to its output, and Anthropic's help center says Claude supports C2PA-based Content Credentials on files to meet EU AI Act Article 50(2). Usually the brand's job is to preserve and extend that record, not to create one from scratch.

A valid credential is not a verdict on truth
How is a Content Credential different from a visible label or a watermark?
A Content Credential differs from a visible label and an invisible watermark in two ways: how it binds to the file and how much information it carries. A visible label is human-readable text, such as "Made with AI" in a caption or overlay. Anyone can read it. But it is not cryptographically bound to the file, so it can be cropped out, mistyped, or put on the wrong asset without any system noticing.
An invisible watermark is what C2PA calls a soft binding. It is embedded in the pixels or the audio signal and survives re-encoding, resizing, and many screenshots. The tradeoff is capacity. A watermark carries only an identifier, not the history. It can tell a verifier that something is there, but it takes a lookup to say what. TrustMark, an open-source watermark from the Content Authenticity Initiative ecosystem, is one example.
The manifest uses a hard binding, which is a cryptographic hash of the asset's content. The hash is what makes tampering detectable, and the manifest carries the full record of origin, edits, and AI use. Its weakness is fragility. Because the manifest travels as embedded metadata, a platform upload, a format conversion, or a screenshot can strip it. A re-encode also breaks the hash, even when the metadata survives.
None of the three works alone. That's why the C2PA community describes Durable Content Credentials as a mix of three signals: the hard-bound manifest, an invisible watermark, and fingerprinting. If the manifest is stripped, the watermark or fingerprint lets a verifier find it again in a cloud repository. OpenAI takes a similar view and treats metadata credentials and embedded watermarks as complementary signals, not alternatives.
For brands, the practical result is layered disclosure. The manifest gives machines a record they can verify. The watermark gives that record a way back after stripping. The visible label gives readers a disclosure they never have to click to see. The comparison table at the end of this guide shows where each approach holds up and where it breaks.
Where should brands attach Content Credentials in the content workflow?
Brands should attach or extend Content Credentials at three points: generation, edit, and publish. Each point records a different fact, and skipping one leaves a gap that a verifier can see.
1. At generation. The AI tool or provider creates the first manifest, with a c2pa.created action and digitalSourceType set to trainedAlgorithmicMedia. OpenAI already does this, and Adobe Firefly attaches credentials when it creates an image. The brand's task is to keep the original file and record that it has a manifest before anyone edits it.
2. At edit. Every design tool that changes the asset should add its own actions to the manifest and re-sign it. That way the history shows the generated base plus any human retouching, cropping, or compositing. Adobe's tools work this way. Tools without C2PA support usually save a new file with no manifest, and this is the most common place provenance quietly disappears inside a brand's own process.
3. At publish. Before an asset goes live, something has to confirm three things: the credential is still present and valid, the AI-use disclosure matches what the team knows about the asset, and the on-page label will render. If the asset was transformed for delivery, this is also when the team decides whether to serve the original or re-sign the transformed copy.
Brands most often leave the publish step to chance, because it falls between the creative team and the delivery stack. Signing at publish takes a C2PA SDK, such as the open-source c2pa tools from the Content Authenticity Initiative, plus a signing certificate the brand controls. Checking at publish takes reading the manifest and comparing it with the asset record.
A useful rule of thumb: the manifest records what happened to the file, and the brand's asset record holds what the brand decided about it. The two have to agree when the asset publishes. The publish step is where that agreement gets checked, so nobody has to assume it.
Why do Content Credentials get stripped downstream, and how do brands keep disclosure intact?
Content Credentials get stripped downstream because the manifest travels as embedded metadata, and many systems throw metadata away by default. Many social and messaging platforms strip it on upload. Format conversions drop it, and screenshots never had it. Any re-encode, such as converting a PNG to WebP or resizing for a responsive layout, breaks the hard-binding hash even if the metadata block survives.
Image delivery pipelines lose credentials quietly. Most modern websites serve transformed images that have been resized, recompressed, or converted to a newer format. Those transforms help performance, and they can drop embedded metadata. Until they've tested it, brands should assume that copies served from an image CDN may not carry the original manifest, including the CDN in front of their own content platform.
Three practices keep disclosure intact. First, use Durable Content Credentials, which combine the manifest with an invisible watermark and fingerprinting so a verifier can recover a stripped manifest from a cloud repository. Second, keep a manifest reference in the asset record itself: whether a credential was present at intake, who issued it, which digitalSourceType it declared, and where the original file lives. That record outlasts every copy of the file. Third, where provenance matters most, serve the original signed file or re-sign the transformed copy. Don't count on a transform to pass metadata through.
The structured record matters more than it first seems. When a platform strips the manifest from a post, the brand's record still knows the asset was AI-generated, and the brand's website can still render a label from that field. If disclosure depends only on file metadata, the brand doesn't control it.
This is the counter-intuitive part of adopting C2PA. The strongest provenance strategy isn't the one that puts the most signatures on the file. It's the one where the brand's system of record holds the disclosure as data and treats the file credential as one of several outputs from that data.
How does C2PA disclosure reach audiences on platforms and websites?
C2PA disclosure reaches audiences in three ways: verification tools that read the manifest, platform badges that show it automatically, and on-page labels the brand renders itself.
Verification tools go deepest. Anyone can upload a file to contentcredentials.org/verify to inspect its manifest, and LinkedIn points users there. Few readers verify by hand, but journalists, regulators, and partners do. The tool shows whether the credential is intact and who signed it.
Platform badges are where most people encounter C2PA. LinkedIn adopted C2PA on May 15, 2024, and shows a Cr icon on images and video that carry signed credentials. Clicking it shows an assertion about the media (whether AI was used to generate or edit it), the app or device used, who created it, who issued the credential, and when. On Sponsored Content, the icon appears only on uploaded single-image or single-video ads, so brands running other ad formats shouldn't expect the badge. TikTok describes itself as the first video-sharing platform to implement C2PA, and it auto-labels AI-generated content uploaded from other platforms by reading Content Credentials. YouTube can use C2PA metadata to trigger its AI label.
Badges only work if the manifest survives upload. If a brand exports a re-encoded file from a tool without C2PA support, the post goes out with no badge, even when the original was signed.
On-page labels are the channel the brand fully controls. A caption, overlay, or accessible text label rendered from a structured AI-use field reaches every reader on the brand's website or app, whether or not they click anything. It keeps working even after a CDN or a partner system has stripped the file credential.
The real goal is consistency across all three. The badge, the verification result, and the on-page label should say the same thing about the same asset. That only happens when the label and the credential come from one source of truth.
EU AI Act Article 50(2) applies from 2 August 2026
How should brands model provenance fields and approval steps for AI-generated assets?
Brands should model provenance as structured fields on every asset record and make publishing wait on an approval step that checks those fields. Free-text captions and tribal knowledge lead to inconsistent labels. Typed fields with fixed options produce the same label every time.
A practical provenance model needs only a few fields:
- AI use: a controlled list of values (fully AI-generated, AI-edited, no AI) that maps to the C2PA digitalSourceType.
- Generator: the tool or provider that made the asset.
- Credential status: whether a manifest was present at intake, whether it validated, and who issued it.
- Manifest reference: a pointer to the original signed file or a cloud-hosted manifest.
- Disclosure label: the exact text audiences see, derived from the AI-use value instead of typed by hand.
- Approval record: who confirmed the disclosure and when.
Controlled values matter because platforms and regulators care about distinctions that editors tend to blur. An image produced entirely by a model and an image with an AI-filled background are different claims. When the AI-use field only allows defined options, the on-page label, the C2PA action, and the reports all carry the same distinction.
The approval step closes the loop. Before an asset with an AI-use value can publish, a reviewer or an automated check confirms the credential status, the label, and that the two match. An asset marked AI-generated with no manifest and no label shouldn't publish silently. It should stop and wait for a decision.
Once the fields exist, reporting is easy. A team can see which AI-generated assets went live in a quarter, which ones carried valid credentials, and which published without a disclosure label. Those are the answers a compliance review or a platform dispute will ask for.
In Sanity, these fields sit in the Schema next to the content that uses the asset. The same AI-use value that drives an on-page label is available to every channel that queries the Content Lake, and the Schema keeps the controlled values consistent across teams and campaigns.
Where does Sanity fit in a C2PA disclosure workflow?
Sanity is the system of record and the place disclosure rules get enforced in a C2PA workflow. The signing itself happens in a dedicated C2PA SDK that Sanity calls. Sanity is the AI Content Operating System, the intelligent backend for companies building AI content operations at scale. In a provenance workflow, that comes down to three jobs: holding the provenance fields, running the checks, and answering audit questions.
Model your business. Provenance and AI-use fields sit in the Schema alongside the campaign, article, or product the asset belongs to. The disclosure is structured content, not a caption, so it reaches the website, the app, and partner feeds in the same form.
Automate everything. Functions are small, single-purpose pieces of code that run on Sanity's cloud infrastructure and react to changes in content. They can read and write the dataset, traverse references, and call external services. A Media Library Asset function can react when an asset changes, call an external c2pa SDK to read the manifest, and write the credential status and digitalSourceType back to fields. Publishing a document fires an update on the published document, so a document function listening for that update can check that every AI-generated asset has a label and then stamp the record. A Blueprint in the repository describes the Functions, so the check is versioned code rather than a manual habit. Sanity doesn't sign or issue Content Credentials. Any re-signing is done by the external SDK the Function calls.
Approval can run through Workflows, which is in beta and opt-in. Workflows are written in TypeScript and deployed like the rest of your code. An agent can check a submission against your rules and either move it forward or send it back.
Audits become queries. Because the process leaves a trail in the content repository, "which AI-generated assets published without a disclosure label" is one GROQ query instead of a spreadsheet exercise. Other CMSes bolt on AI, while Sanity is built for it. In practice, the disclosure rule lives in the same structured system as the content it governs.
Ways to disclose AI-generated media: manifest, label, watermark, or a Sanity record
| Feature | Sanity | C2PA manifest only | Visible label only | Invisible watermark only |
|---|---|---|---|---|
| What the disclosure carries | Typed AI-use, generator, credential status, and manifest reference fields in the Schema, which every channel can query from the Content Lake | The full signed history: origin, edits, and AI use, recorded in c2pa.actions with digitalSourceType set to trainedAlgorithmicMedia | Human-readable text such as 'Made with AI', with no history, no issuer, and nothing behind it that a machine can verify | An identifier embedded in the pixels or audio. It carries no history itself and needs a lookup to return provenance details |
| Survives metadata stripping and re-encoding | The record survives: the field persists when a platform strips the file. Pair it with watermarking so the file credential can be recovered too | No. Platform uploads, format conversions, and screenshots strip it, and any re-encode breaks the hard-binding hash | Only if it's baked into the pixels. Caption labels disappear when the image is shared without the page around it | Yes. A soft binding is designed to survive re-encoding and resizing, and it can point a verifier to a stored manifest |
| Tamper evidence | Changes to provenance fields leave a trail in the content repository. Cryptographic tamper evidence still comes from the signed manifest | Strong. The manifest is signed and bound by a cryptographic hash, so verifiers can detect edits and check whether the signer is on a trust list | None. The text can be removed, changed, or put on the wrong asset without any system noticing | Limited. It shows an identifier is present but doesn't on its own prove the content is unchanged since signing |
| Triggers platform AI badges | Indirectly. Functions can check that the published asset still carries its manifest so LinkedIn, TikTok, and YouTube can read it | Yes, when intact. LinkedIn shows a Cr icon, TikTok auto-labels, and YouTube can trigger its AI label from C2PA metadata | No. Platforms read machine-readable metadata, not caption text, so a visible label alone triggers no badge | Not directly. Its value is letting verifiers find a stripped manifest, which is its role in Durable Content Credentials |
| Consistent labeling across assets | Controlled AI-use values in the Schema drive the label text, so every AI-generated asset gets the same disclosure wording | Depends on every tool in the chain supporting C2PA. One editor without support can break the history for that asset | Depends on editors remembering the label and typing it correctly, which drifts across teams, agencies, and campaigns | Consistent if one pipeline step applies it, but readers can't see it, so it never discloses anything to them on its own |
| Answering audit questions | 'Which AI-generated assets published without a disclosure label' is one GROQ query. Approval can run through Workflows (beta) | Files have to be inspected one at a time with a verifier such as contentcredentials.org/verify, which doesn't scale to a whole catalog | Someone has to review published pages by hand to find labels that are missing, outdated, or worded inconsistently across channels | Files have to be scanned for the watermark. A detection confirms it's there but not what the brand decided to disclose |