presentsUNAIVERSE
The UN AI timeline

    Learn AI /Generative Media /Watermarking and content provenance

    Watermarking and content provenance

    Also called: content credentials, C2PA, provenance.

    Detail

    In five seconds

    What it is

    Content provenance records claims about origin and edits, while watermarking embeds a detectable signal that may indicate a source or how media was made.

    In plain words

    Think of this as a way to carry evidence about where a photo, video or audio file came from and what happened to it later. Some systems attach a signed history next to the file. Others hide a marker inside the file. Both can help with checking, but neither guarantees truth.

    An everyday example

    A second-hand bicycle may come with a service booklet containing signed entries for assembly and repairs. It may also have a small maker's mark stamped into the frame. The booklet resembles provenance: it records claims about where the bicycle came from and what was changed. The stamped mark resembles a watermark: it is a detectable signal associated with a source. A missing booklet does not prove the bicycle is counterfeit, a genuine booklet does not guarantee every entry is accurate, and the maker's mark does not reveal the repair history.

    Where the story stops working

    Real provenance systems use cryptography and software verification, not handwritten logs, and watermarks may be invisible rather than faintly visible. Also, a library stamp usually points to one institution, while media provenance can involve many tools and organisations across a chain. A watermark detector can also make mistakes, which a simple stamp analogy does not show well.

    Remember thisProvenance carries verifiable claims about origin and edits; watermarking embeds a detectable signal. Either may be missing, stripped or incomplete, and neither alone proves authenticity.

    This term is contested

    Some details are unsettled in practice, especially how effective watermarking is across different media and transformations, and how strongly provenance standards should be treated as evidence when adoption is uneven.

    Understand it

    Understand it

    Why it matters

    For policy, communications and leadership, the key question is often not 'Can this be generated?' but 'What evidence do we have about origin and handling?' Provenance records and watermarks can improve traceability, support platform rules, and help teams assess risk before sharing, quoting or acting on media. They also matter for procurement and governance: different tools preserve different evidence, and some evidence disappears during reposting or editing. A missing credential does not prove deception, and a present credential does not prove the content is accurate in context. Good decisions still require corroboration, source checks and human judgement.

    How provenance records and watermarks support authenticity checks

    After viewing this, a reader should be able to explain the difference between signed provenance records, content credentials, the C2PA standard and watermarks, and why none of them alone proves truth.

    1. 01

      1. Create or capture media

      An image, audio clip or video is made by a camera, editing tool or generative system. At this point, a tool may prepare claims about origin, date, device or model.

      Draw a file card on the left with icons for camera, image editor and AI generator feeding into it. Add a small note icon labelled 'claims'.

    2. 02

      2. Record provenance

      Compatible tools can attach signed metadata that records who or what created the file and what edits happened. Content credentials are one user-facing way to present this information, often using the C2PA standard underneath.

      Show a metadata document attaching to the file card with a lock and signature badge. Add a caption: 'signed provenance / credentials' and a small standard tag 'C2PA'.

    3. 03

      3. Embed watermark

      A system may also place a detectable pattern in the pixels, audio or video signal. This can survive some copying even if metadata is stripped, but it usually carries less detail than a full provenance record.

      Overlay a subtle dot-grid or wave pattern inside the file card. Add a parallel lane labelled 'watermark in content'.

    4. 04

      4. Share and transform

      The file is posted, downloaded, cropped, recompressed or edited. Some platforms preserve metadata; others remove it. Watermarks may survive some changes and fail under others.

      Move the file through boxes labelled 'post', 'download', 'crop', 'recompress'. Show one branch where metadata tag fades, while watermark pattern remains faintly visible.

    5. 05

      5. Verify and interpret

      A viewer or software checker reads credentials if present and tries watermark detection if relevant. The result may show an intact history, only partial clues, or no trustworthy evidence. Human review is still needed.

      End with three outcome panels: green 'intact signed history', amber 'partial source clues', grey 'no reliable provenance found'. Add a person icon reviewing the result.

    A left-to-right flow with one media file moving through creation, optional credential signing, optional watermark embedding, distribution and verification. Two parallel labels show that provenance records travel mainly as metadata, while watermarks travel inside the media signal. The final panel shows three possible outcomes: intact chain, partial evidence, or no reliable evidence.

    Hover or tap each stage to highlight what information is added, what may be lost during editing or upload, and what a verifier can still check at the end.

    Describe this without the diagram

    The diagram is a five-step flow from left to right. Step 1 shows media being created by a camera, editor or AI system, with possible claims about origin. Step 2 shows signed metadata attached to the file; this is the provenance record, often displayed as content credentials and commonly implemented using the C2PA standard. Step 3 shows a watermark embedded inside the media itself as a separate path. Step 4 shows sharing and editing, where metadata may be removed and watermarks may weaken but sometimes remain. Step 5 shows verification producing one of three outcomes: intact signed history, partial clues, or no reliable evidence. The main message is that provenance records and watermarks are different tools and neither alone proves that content is true.

    With reduced motion: A static five-column layout shows the same process without animation. Solid lines connect creation to provenance recording and watermark embedding, then to sharing and verification. Colour and labels indicate what travels as metadata and what travels inside the media signal.

    These terms are related, but they are not identical. Content provenance means information about a file’s origin and history: who created it, with what tool, and what edits happened later. Content credentials are a practical way to package and display some of that information. C2PA is one technical standard used to record and verify such claims across compatible tools. Watermarking usually means adding a signal to the content itself, such as pixels, audio patterns or text markers, so a system may detect it later.

    What goes in is a piece of media, such as an image, audio clip or video, plus claims about it. Those claims might say “captured on this device”, “edited in this application”, or “generated by an AI system”. What happens next depends on the method. A credentials system can attach signed metadata, meaning extra data protected by cryptography so later changes can be detected. A watermarking system changes the content or adds hidden patterns so software can try to recognise the source even if metadata is missing.

    What comes out is not automatic proof of truth. It is evidence about source and change history that may support verification. If the chain is intact, a viewer or tool may see that a photo was taken on one device, cropped in another app and then exported. If a hidden watermark is present, a detector may estimate that the file likely came from a certain model or platform.

    These ideas are often confused with transparency and with deepfake detection. Transparency is broader: it includes disclosure, documentation and clear communication, not only technical markers in files. Deepfake detection tries to judge whether content may be synthetic or manipulated, often from patterns in the content itself. Provenance and watermarking instead try to preserve or signal origin information from creation onward. They can help each other, but neither one solves authenticity on its own.

    How “A service booklet and a maker's mark” maps onto the real thing

    In the storyIn the AI system
    signed handling logcontent provenance recorded as signed metadata or credentials
    entries for scan, crop and label changesclaims about creation and later edits
    faint pattern on public copieswatermark embedded in the media
    missing log but visible stamp cluewatermark may help when metadata is removed

    A worked example

    An information management officer at a regional hub is checking a photograph submitted by an implementing partner for a monitoring report due for upload to the portal by close of business.

    Starting from: The submitted JPEG, the original camera file and the proposed report caption

    1. An approved checker reads a signed manifest claiming camera capture followed by cropping and a tonal adjustment.
    2. The signature validates and the file matches the signed manifest. This shows that the recorded history has not been altered since signing, not that the depicted place, time or activity is correct.
    3. A portal-ready copy has lost its credentials during export. A separate detector finds no recognised generative-media watermark, but neither absence establishes authenticity.
    4. The officer compares the original file, the recorded transformations and the proposed caption, then asks the field focal point to confirm the context.

    Result: The photograph is used only after the provenance record and contextual checks are consistent. The officer records the checks in a note for the file and retains the credential-bearing original.

    In a United Nations workplace

    In a Headquarters department, a training officer is finalising a statement for delivery that introduces an online course. The statement and its accompanying slide need front-office clearance by close of business, but the vendor-supplied image carries content credentials and a faint watermark.

    The credentials may contain signed claims that the image was generated and later cropped, while the watermark may provide a separate signal about its source. Verifying the signature shows whether the manifest has changed since signing. It does not establish that every claim in the manifest is true or that the image is suitable for use.

    Before you try thisDo not upload the uncleared statement, internal slides or other non-public material to a public provenance checker. The service may retain or expose them. Use an approved tool, confirm the image licence and intended use, and do not treat missing credentials or an undetected watermark as proof that media is authentic.

    Where you may hear it

    • A vendor demonstration to a technology governance board
    • Front-office clearance of a statement using externally supplied images
    • A communications review before material is uploaded to the portal
    • A procurement meeting for a generative-media service

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • If a file has content credentials, the content must be true.

      Credentials can show who or what produced a file and whether the recorded history stayed intact. They do not confirm that the scene, caption or surrounding claim is accurate. A genuine photo can still be miscaptioned or shown out of context.

    • Watermarking and provenance are the same thing.

      They solve related but different problems. Provenance usually records a chain of source and edits, often in signed metadata. Watermarking places a detectable marker inside the content itself. A watermark may survive when metadata is removed, but it usually carries less detailed history.

    • If no watermark or credentials are found, the file is fake.

      Many ordinary files have no provenance data because the camera, software or platform did not create or preserve it. Absence of evidence is not evidence of fabrication. It means you need other checks.

    • C2PA is a detector that can identify all AI-generated media.

      C2PA is mainly a standard for recording and verifying claims about a file’s origin and edits across compatible tools. It is not a universal detector, and it depends on adoption, preservation and honest use of the system.

    Quick check

    Which statement best captures the relationship among these terms?

    Go deeper

    Go deeper

    The simple version

    Think of this as a way to carry evidence about where a photo, video or audio file came from and what happened to it later. Some systems attach a signed history next to the file. Others hide a marker inside the file. Both can help with checking, but neither guarantees truth.

    The technical version

    Content provenance refers to machine-readable assertions about a digital asset’s origin, capture, generation and transformation history. Content credentials are a user-facing presentation of such assertions, often implemented through cryptographically signed manifests. C2PA, the Coalition for Content Provenance and Authenticity specification, defines a common way for producers, editors and distributors to bind assertions, ingredients and actions to an asset and enable later verification of integrity and lineage in compatible systems. Watermarking is different: it encodes a detectable signal into the media itself or inserts identifiable patterns in generated outputs. Robust watermarking aims to remain detectable after ordinary transformations such as resizing or compression, though robustness, payload size and imperceptibility trade off against each other. Provenance metadata can be stripped; watermarks can be damaged or produce uncertain detection results. In practice, stronger authenticity workflows combine signed provenance, platform preservation, policy rules and independent verification rather than relying on one mechanism alone.

    Reviewed external resources

    Where this came from

    Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

    The one thing to remember

    Provenance carries verifiable claims about origin and edits; watermarking embeds a detectable signal. Either may be missing, stripped or incomplete, and neither alone proves authenticity.