Blendpipe homeSend a brief
Menu

C2PA Content Credentials for brand video: what to attach and what survives

C2PA Content Credentials on video ads: what a manifest records, where it survives from edit to upload, what platforms do with it and when to sign masters.

Reader question
Should a brand attach C2PA Content Credentials to its video ads, and what happens to them between the edit suite and the platform?
A canvas tote bag hangs from the back of a cafe chair on a sunlit sidewalk, and a drawn marigold luggage tag tied to its strap is slipping loose in the breeze.

Yes, a brand should keep or attach C2PA Content Credentials on its video masters when its tools can write them, but it should expect much of that record to be stripped or rewritten before a viewer sees the ad. A Content Credential is a signed manifest that describes how a file was made, including any generative AI step, and it survives only as far as each editor, encoder and platform preserves or re-signs it. YouTube, TikTok and LinkedIn read credentials on upload, and a YouTube AI label applied because of C2PA metadata cannot be adjusted by the uploader. So decide what each file says before it ships, test it on the platform, and hand over a provenance log that does not depend on the embedded record.

This guide reads the C2PA Technical Specification (version 2.5, dated October 2026), the C2PA explainer and implementation guidance, the IPTC’s FAQ for publishers, and each platform’s own help pages, all checked in November 2026. It covers what a manifest records, what happens to it at each stage of an ad pipeline, what each platform does with it, and the three decisions a brand owns: whether to sign masters, which edits to record and what goes in the handover. Whether a given ad needs an AI label at all is a separate question, answered in AI disclosure labels on video ads.

What a C2PA Content Credential records

A Content Credential is a C2PA manifest: a set of statements about a file, bundled into a claim that a tool signs with a certificate and binds to the file’s exact bytes with a hash. It records how the file began, which actions were applied to it and which other files went into it.

“Content Credentials” is the public name; C2PA is the specification underneath. The IPTC’s FAQ (version 1.0, last updated July 8, 2026) puts the division of labor plainly: “C2PA writes the standard, CAI promotes it and builds tools for it.” The small “cr” pin you see on some sites is the Content Credentials icon.

The parts of a manifest

The C2PA Technical Specification 2.5 defines four parts that matter to a producer:

  • Assertions. Each is “a data structure which represents a statement” made by the signer: the actions taken, the ingredients used, a thumbnail, metadata.
  • The claim. “A digitally signed and tamper-evident data structure that references a set of assertions.”
  • The claim signature. The digital signature on the claim, made with the signer’s private key.
  • The manifest store. “A collection of C2PA Manifests that can either be embedded into an asset or be external to its asset.”

Every standard manifest must contain exactly one hard binding, which the specification defines as “one or more cryptographic hashes that uniquely identifies either the entire asset or a portion thereof.” That hash is what makes the record tamper-evident. Change one frame, or re-encode the file without re-signing it, and the hash no longer matches the content.

The actions an ad pipeline writes

The actions assertion lists what was done to the file. Version 2.5 defines a fixed vocabulary of action names, and a tool may add its own (the specification’s example is com.fabrikam.gaussianBlur). These are the standard actions a commercial is most likely to carry:

Action What the specification says it means Where it shows up in ad work
c2pa.created The asset was first created or generated, whether by a person, a camera or AI Camera originals, generated clips, a new edit
c2pa.opened An existing asset was opened as the parent ingredient Opening the hero master to make a cutdown
c2pa.placed One or more component ingredients were added Plates, renders or generated clips dropped into a timeline
c2pa.adjustedColor Changes to tone, saturation and similar The grade
c2pa.cropped, c2pa.resized Areas cropped out; dimensions or file size changed Reframing 16:9 to 9:16 or 4:5
c2pa.trimmed Time removed from the start or end Cutting a 30-second spot to 15
c2pa.addedText Visible text inserted, such as a caption Supers and burned-in captions
c2pa.dubbed, c2pa.translated Audio changed; language changed A Spanish version
c2pa.transcoded One encoding converted to another, including resolution and bitrate changes Delivery encodes
c2pa.repackaged Container changed without transcoding A rewrap
c2pa.published The asset was released to a wider audience Release or upload
c2pa.reviewed Content was read and assessed without modification QC, legal or client review

Two details from the same document are useful. c2pa.color_adjustments is now deprecated in favor of c2pa.adjustedColor, and version 2.5 added c2pa.modified, c2pa.rendered and c2pa.reviewed. The specification also says that a manifest containing only c2pa.published, c2pa.transcoded and c2pa.repackaged actions signals that “no editorial changes” happened between the source and this file. A properly signed delivery encode can therefore tell a reader that nothing was edited after the approved master.

How AI generation is recorded

The creation action must carry a digitalSourceType value describing how the asset began. The specification’s own example is a video generated from a text prompt, which carries trainedAlgorithmicMedia. The C2PA explainer says that when an AI system performs an action, it “is clearly identified as such” through the digitalSourceType field.

The values come from the IPTC Digital Source Type vocabulary. These are the ones an ad’s ingredients are most likely to carry:

Value IPTC label Typical ad ingredient
digitalCapture Digital capture sampled from real life Camera footage
humanEdits Human-edited media A conventional edit or composite
digitalCreation Digital creation (non-generative tools) Hand-made animation, 3D renders
trainedAlgorithmicMedia Created using Generative AI A generated clip or still
compositeWithTrainedAlgorithmicMedia Edited using Generative AI, such as inpainting or outpainting AI-assisted cleanup or extension of a real plate
compositeSynthetic Composite including generative AI elements A finished spot with at least one generated element

Version 2.4 (April 2026) added an AI Disclosure assertion (c2pa.ai-disclosure) that can name the model type and a humanOversightLevel such as prompt_guided or human_validated. Version 2.5 lets an editing tool attach that assertion to an ingredient that came from a tool without C2PA support. Both are recent, so confirm your tools write them before anyone relies on them.

Ingredients carry their own history. The explainer describes “a tree of provenance, much like a family tree,” so a three-second generated sky placed in a 30-second spot can appear in the finished master’s credential if the editing tool recorded it as an ingredient. How a platform weighs one AI ingredient inside a mostly filmed spot is not published in detail, which is why the testing step later in this guide matters.

What Content Credentials prove and what they do not

A Content Credential proves that a signed record exists and that the file has not changed since it was signed. It does not prove that the content is true, and it does not detect AI in files that carry no record.

The C2PA explainer says credentials “do not provide value judgments about whether a given set of provenance data is ‘true’,” only whether the record is well formed, untampered and signed by a trusted signer. The IPTC FAQ goes further for audiences: credentials “make no statement about the factual accuracy of the content itself,” and “C2PA is not a deepfake detector.” Software that reads C2PA “simply extract[s] the data that was put there by its creator or publisher.”

For an advertiser, that has three practical consequences:

  1. A credential does not substantiate a claim. A manifest showing camera-original footage of a bottle says nothing about whether the bottle keeps coffee hot for twelve hours. Claim substantiation stays with the brand and its counsel.
  2. A credential is not a license. The explainer says Content Credentials, unlike DRM, do not “restrict user access or impose usage limitations.” Usage terms for music, talent and stock still live in contracts and the rights sheet described on rights and usage.
  3. Trust depends on who signed. The explainer describes a trust list for accredited implementations, and the IPTC FAQ says the C2PA trust ecosystem now limits signing of the manifest itself to tools (hardware and software) that pass the C2PA Conformance Program. A person or company adds its own identity through a separate signed identity assertion, defined by the Creator Assertions Working Group (CAWG).

Following a spot from camera to feed

Credentials are written at capture or generation, kept or rewritten in the edit, re-signed at export, and most often lost in transcoding and distribution. The diagram follows one spot through the stages where that happens.

Where a Content Credential is written, kept or lost in an ad pipelineSeven stages from capture to viewer. Capture or generation and the export of a master write a manifest. The edit and each version keep it only if the tool reads and re-signs it. Transcoding, ad servers and content delivery networks often strip it. Platforms read it on upload and may apply a label. Viewers usually receive a platform rendition without the manifest. Stage What happens to the credential 1 Camera or generator Written: c2pa.created plus source type 2 Edit, composite, grade Kept only if each tool reads and re-signs 3 Export master (MP4, MOV) New signed manifest; ingredients linked 4 Cutdowns and reframes Each version needs its own manifest 5 Transcode, ad server, CDN Often stripped unless set to keep it 6 Platform upload Read at ingest; a label may follow 7 What the viewer receives A rendition, often without the manifest written depends on the tool often lost read by the platform
Typical stages for a mixed-media spot; your pipeline may add or skip stages. / Sources: C2PA Technical Specification 2.5; C2PA Guidance for Implementers 2.2; IPTC C2PA FAQs v1.0; YouTube, TikTok and LinkedIn help pages, checked November 2026.

Capture and generation

The IPTC FAQ’s list of shipping products with C2PA names still cameras from Leica, Nikon, Canon and Sony, and one video camcorder, Sony’s PXW-Z300. Unless your footage comes from a camera that writes C2PA, the chain starts at the edit, not on set. Check your camera’s documentation rather than assuming.

Generators are further along. The same FAQ says “OpenAI, Google, Adobe and Microsoft all embed C2PA Content Credentials in their AI-generated content.” A generated clip therefore often arrives in the edit with a manifest that already says trainedAlgorithmicMedia, whether or not anyone on the team looked at it.

Edit, composite and grade

The edit is where the chain is kept or broken. Adobe’s help page on Content Credentials in Premiere and Media Encoder (last updated September 9, 2026; read through a Wayback Machine capture of November 4, 2026, because the live page refused automated requests) describes the workflow:

  • A setting, Validate Content Credentials on import, checks incoming media; Scan for Content Credentials in the Project panel checks assets by hand.
  • When a sequence contains Adobe AI-generated assets, Premiere and Media Encoder say that credentials will be included in the export for supported formats: MP4, QuickTime MOV, AVI, WAV, MP3, JPEG, PNG and TIFF.
  • In the Export window, Apply Content Credentials preferences signs the file with your verified name and linked accounts, and the preferences page can request that generative AI models not train on your content.
  • Adobe cautions that publishing destinations may not support the credentials it writes.

A composite or finishing tool that does not read C2PA produces a render with no manifest. The specification’s own guidance covers this case: a broadcaster example uses an invisible watermark to carry provenance through “non-C2PA capable” stages, then retrieves the earlier manifest afterward. If your pipeline has no such watermark step, the gap belongs in the provenance log.

Export and masters

The specification embeds manifests in ISO BMFF files, which include MP4 and QuickTime MOV, and in streaming formats such as CMAF, plus caption formats WebVTT and TTML2. MXF does not appear in the version 2.5 list of embeddable formats. If a broadcast spec asks for MXF, the options are an external manifest, which the specification allows when “embedding may not be technically possible,” or the provenance log alone.

Ask whether your signing tool adds a trusted time-stamp. The C2PA implementation guidance explains that without one, consumers cannot trust that an asset signed with a now-expired credential was signed before it expired. A spot that may be recut or re-aired years later benefits from the time-stamp.

Versions, transcodes and ad servers

Every cutdown, reframe and language version is a new file, so each needs its own manifest: the master opened as the parent, then trimmed, cropped or dubbed. Every encode an ad server or content delivery network makes is a new binary too. The C2PA guidance says metadata “may be routinely removed or corrupted by legacy or non-C2PA capable platforms during distribution,” which is “common” on social platforms that display renditions.

The IPTC FAQ gives publishers two fixes that translate to ad delivery. “Most CDN tools can be configured not to strip metadata,” and when a system makes several variants for different screens, “all variants” should be signed. Where stripping cannot be avoided, it recommends soft bindings: fingerprints or invisible watermarks that let a reader find the manifest again in a repository.

Upload and playback

Platforms read the manifest at ingest; viewers usually receive the platform’s own rendition. The IPTC FAQ notes that LinkedIn shows the “cr” pin on signed images, but “the image displayed by LinkedIn has the C2PA metadata stripped out,” so a viewer cannot load that copy into a validator. Soft bindings help here, with a limit the FAQ states: fingerprint and watermark matching is statistical and has false positives and false negatives. The FAQ adds that watermarking and C2PA can be applied to the same content as separate signals, and that the EU AI Act Code of Practice recommends that pairing for AI-generated content.

What YouTube, TikTok, Meta, LinkedIn and Google do with credentials

YouTube, TikTok and LinkedIn read C2PA credentials on upload and show something to viewers; Meta labels content when it detects industry signals; Google says its ad systems have started to use C2PA. Of the pages below, only LinkedIn’s says what happens to credentials on ads.

Platform What it reads What viewers see Can the uploader change it?
YouTube C2PA metadata in the upload An AI label on the player or in the expanded description; “Captured with a camera” in How this content was made, for qualifying files Not for labels applied because of C2PA metadata
TikTok Content Credentials on images and videos uploaded from other platforms (since May 9, 2024) An automatic AI-generated label Not stated
Meta “Industry standard indicators” plus self-disclosure “AI info” label; for content only edited with AI, the label sits in the post menu Not stated
LinkedIn C2PA Content Credentials on images and videos The C2PA icon; a click shows the credential Not stated; rolling out gradually
Google ad systems C2PA metadata, “starting” as of September 17, 2024 Not described Not described

All checked November 22, 2026. Platform behavior changes, so recheck before each campaign.

YouTube: labels that stick

YouTube’s disclosure page lists three cases in which it “may automatically apply an AI label”: content made with YouTube’s own generative AI tools, “Content that contains C2PA metadata,” and content its internal systems detect as AI generated or altered. Creators can usually change a disclosure, but labels on content made with YouTube’s tools, content containing C2PA metadata and content labeled after manual review cannot be adjusted. A separate page on How this content was made says YouTube carries forward disclosures from Content Credentials 2.1 or higher that indicate the entire video is made with AI, often with an “Info from” attribution to the signer.

The Captured with a camera disclosure works the other way. It requires capture with tools that have built-in C2PA support at version 2.1 or higher, and the content “must also not have edits to sound or visuals.” A finished commercial is edited by definition, so it will rarely qualify. YouTube adds that a missing label “doesn’t mean the content has modified audio or visuals.”

TikTok

In its May 9, 2024 announcement, TikTok said it was “launching the ability to read Content Credentials,” starting with images and videos, and would automatically label AI-generated content uploaded from other platforms when it carries them. It also said it would attach credentials to TikTok content and that they “will remain on content when downloaded.” The announcement does not say whether ad uploads go through the same check.

Meta

Meta’s labeling post, first published April 5, 2024, says labels rely on “industry standard indicators” and self-disclosure. The label became “AI info” on July 1, 2024, and since a September 12, 2024 update, content Meta detects as only edited with AI shows the label in the post menu. An earlier February 6, 2024 post names the C2PA and IPTC standards and says Meta could not yet detect those signals in video and audio from other companies at the same scale. Check Meta’s current help pages before you count on video detection.

LinkedIn

LinkedIn’s help page says images and videos signed with C2PA Content Credentials carry the C2PA icon, and a click displays the credential: whether AI was used, the app or device, the issuer and the issue date. It is the only platform page here that mentions ads: for Sponsored Content, the icon appears on single image or video ads “when the image or video is uploaded,” including ads built in Campaign Manager. LinkedIn says the feature is rolling out gradually.

Google’s ad systems

Google’s September 17, 2024 post says “Our ad systems are starting to integrate C2PA metadata” and that the goal is to “use C2PA signals to inform how we enforce key policies.” It does not say what advertisers will see. Treat it as a direction of travel, not a rule you can plan around.

Should a brand sign its masters?

Sign when the record helps you and you can keep it accurate: when the spot contains generated material you will disclose anyway, when a platform or buyer asks for provenance, or when you want copies of your ad traceable to you. Hold back, or re-sign from a clean slate, when the record would expose private details or describe the spot inaccurately.

Reasons to sign

  • The record matches the disclosure. If the spot contains a generated shot that needs a label under platform rules, a credential that says so is consistent with the label you will apply.
  • Tampering becomes visible. The IPTC FAQ notes that “if signed content is modified in any way, the content’s digital signature will no longer be valid.” An altered copy of your ad reposted elsewhere fails validation.
  • Copies trace back to you. With a CAWG identity assertion, IPTC says, provenance travels with the content so it “can be traced back to the organisation that published it.”

Costs and constraints

  • Certificates. For news publishers, the IPTC FAQ puts an identity certificate at “hundreds to low thousands of dollars or euros per year,” depending on the certificate authority. It says most of the cost is integration work, not the certificate.
  • Conformant tools only. The manifest itself is signed by a conformant tool, so your options depend on what your editing, encoding and delivery tools support.
  • Privacy. Manifests can carry names, linked accounts and thumbnails of ingredients. For an unreleased product, an ingredient thumbnail can expose a frame you meant to keep internal.
  • Labels you cannot reverse. On YouTube, a label applied because of C2PA metadata cannot be adjusted, so the manifest has to be right before upload.

Whose identity goes on the file

Decide this in the contract, not at export. The C2PA explainer says the signer of a Content Credential is the holder of the certificate the signing tool uses, and only the assertions the signer created are attributed to it. If a studio or agency exports with its own linked identity, copies of the ad trace back to the studio, not the brand.

Three workable arrangements exist. The production company signs working files and the brand signs final deliverables with its own identity assertion; the brand signs nothing and keeps the log; or the agency signs on the brand’s behalf under a written instruction. Whichever you choose, name it in the production plan so every editor exports the same way.

Clean-slate signing

The specification allows redaction of assertions, but the IPTC FAQ says that part “is not yet well implemented across all tools.” For news publishers it recommends “clean slate stamping”: strip all C2PA metadata at publish time, then add a manifest containing only what the publisher is willing to attest, with its identity assertion. A brand can borrow the approach for final deliverables, with one condition: the new manifest must still record every generative AI step the platform rules or law require you to disclose.

Stripping is not a disclosure plan

The IPTC FAQ says C2PA metadata “is removable by design,” because many legitimate tools remove metadata. Removing it does not remove a disclosure duty under a platform’s rules or the law, and YouTube’s list of label triggers includes its own detection systems, not only C2PA metadata. Whether a particular spot must be labeled is a decision for the brand’s counsel, using the platform rules covered in AI disclosure labels on video ads.

SituationSuggested decisionWhy
Spot contains generated shots that will be disclosedKeep the tool-written credentials and sign the masterThe record and the label say the same thing
Only AI-assisted cleanup or rotoscoping, finished by a personRead the manifest, then test an unlisted upload before releasePlatforms do not publish how they weigh edit-level AI signals
Unreleased product or confidential talent details in ingredientsClean-slate sign the deliverables; keep full history in the logIngredient thumbnails and identity fields travel with the file
Broadcast master in MXFExternal manifest or provenance logMXF is not in the 2.5 list of embeddable formats
Ad server or CDN re-encodes every fileAsk whether it keeps or re-signs manifests; otherwise rely on the logEach rendition is a new binary with a new hash
Decision aid, not legal advice; counsel decides disclosure. / Basis: C2PA Technical Specification 2.5, IPTC C2PA FAQs v1.0 and YouTube Help, checked November 2026.

Which edits to record, and at what level of detail

Record every step that changes what the audience sees or hears, and every AI step with its source type; let tools record routine encodes automatically. The embedded manifest carries the actions; the provenance log carries the shot, the frames, the reason and the reviewer.

The two records do different jobs. The manifest is machine-readable and travels with the file until something strips it. The log is a document the client keeps, written for people: producers, counsel, platform reviewers. The specification even notes that the order of actions in a manifest “does not imply the order in which actions were performed,” so sequence and reasoning belong in the log.

Pipeline step What the tool should write Source type if a new asset is created What the provenance log adds
Generated background plate c2pa.created in the generator, c2pa.placed in the edit trainedAlgorithmicMedia Shot, timecode, tool category, reviewer
AI-assisted cleanup of a filmed plate c2pa.edited or a tool-specific action Result: compositeWithTrainedAlgorithmicMedia Frames touched, what was removed, reviewer
AI-assisted rotoscoping for a matte Depends on the tool; often none None if the matte is not a new asset Shot, who finished the edges
Grade c2pa.adjustedColor None Nothing beyond the normal edit record
15-second cutdown c2pa.opened, c2pa.trimmed None Version name from the matrix
9:16 reframe c2pa.cropped or c2pa.resized None Version name
Spanish version c2pa.dubbed, c2pa.translated None Translator and reviewer
Burned-in captions c2pa.addedText None Caption file version
Delivery encode c2pa.transcoded None Encode settings, destination
QC or legal review c2pa.reviewed None Reviewer and date

The rotoscoping row is the easiest to miss. A matte that only separates a filmed hand from its background may never become a separate asset with its own manifest, so the embedded record can be silent about an AI-assisted step. The log is where that step is written down.

Example (fictional). Orlenna is an invented brand, and the Ridge Bottle spot is an illustration, not a client project. The hero spot is 30 seconds at 24 fps, in 8 shots. Two shots use AI-assisted steps: shot 1 (0 to 4 s), where a cleanup removes a power line from the sky, and shot 5 (17 to 20 s), where AI-assisted rotoscoping makes the hand matte and a person finishes it.

Frames to review: shot 1 is 4 s x 24 fps = 96 frames, checked frame by frame; shot 5 is 3 s x 24 fps = 72 frames. Together that is 96 + 72 = 168 frames, which matches the plan’s 7 seconds of AI-assisted work (7 s x 24 fps = 168 frames).

The cleanup result would carry compositeWithTrainedAlgorithmicMedia if the tool writes it; the matte may leave no trace in the manifest. The deliverable matrix is 3 lengths x 4 ratios x 2 languages x 1 call to action = 24 files, and each log entry lists which of the two shots that version contains.

How to test credentials before and after upload

Test the files, not the settings: open each delivered version in a C2PA validator, then upload a private or unlisted copy to each platform and record what appears. Testing before release costs far less than finding a label you cannot adjust after the media buy is live.

  1. On import, turn on credential checking in the editing tool (in Premiere, Validate Content Credentials on import) and note which sources arrive with manifests.
  2. Before export, check the timeline for leftover generated material, such as a concept frame used as a placeholder, that would be recorded as an ingredient.
  3. After export, open the master in a validator. Content Credentials Verify lives at contentcredentials.org/verify, which now redirects to verify.contentauthenticity.org; the IPTC FAQ also names IPTC Origin Verify. Read the signer, the actions, the ingredients and every source type.
  4. Check for exposure. Look at ingredient thumbnails, identity fields and linked accounts before the file leaves the building.
  5. After versioning, validate each cutdown, reframe and language version; they are separate files with separate manifests.
  6. After the ad server or CDN, pull one served rendition per destination and validate it, to learn whether that system keeps, re-signs or strips manifests.
  7. On each platform, upload a private or unlisted test and record any label, its wording and where it appears.
  8. Write the results into the provenance log with the date, because platform behavior changes.

Example (fictional). For Orlenna’s 24 deliverables, validating each file once after export and once after the ad server makes its rendition is 24 x 2 = 48 checks. Suppose each check takes 3 minutes: 48 x 3 = 144 minutes, or 2.4 hours. Add one unlisted test upload per platform, then record every result in the log.

Questions for your ad server, CDN or trafficking vendor

Step 6 usually raises questions that only the vendor can answer. Send them in writing before the campaign:

  • Does your system keep C2PA manifests in files it passes through unchanged?
  • When you re-encode, do you sign a new manifest with a c2pa.transcoded action, or drop the manifest?
  • Can stripping be switched off per account or per campaign?
  • Do you read or add any watermark or fingerprint?
  • Which of your output renditions reach which placements?

The delivery specs checklist is a good place to add the credential checks next to loudness, captions and file specs, so they happen in the same QC pass.

What goes in the provenance log handed over at delivery

The provenance log is the durable record: it survives when the embedded manifest is stripped, it covers steps no tool recorded, and it names the people who reviewed the work. Hand it over with the masters and keep it with the rights sheet.

Field What to write Why it matters
File name The version’s name from the deliverable matrix Ties each entry to one file
Credential status Signed, not signed, external manifest, or stripped downstream Tells the next editor what to expect
Signer The tool that signed and any identity assertion Needed to read validation results
AI steps Shot, timecode, frames, tool category, source type written The disclosure and copyright questions start here
Reviewer The person who checked each AI-assisted output Shows human review happened
Unsigned ingredients Plates, renders and stock that arrived without credentials Explains gaps in the chain
Soft bindings Any watermark a tool added Lets someone recover a stripped record
Platform tests Platform, date, label shown, where it appeared Evidence of what viewers saw
Disclosure decision Label applied or not, and who decided Keeps the decision with the right person

The log also helps with ownership. The US Copyright Office treats human-authored and AI-generated material differently, and a shot-by-shot record of who made what is what counsel needs to decide what to register; Can you copyright an ad made with AI? explains that question. On Blendpipe’s side, the AI practices page sets out how every AI-assisted step is logged and reviewed, and deliverables shows where the log sits in the handover package.

Common failure points and what they mean

Most credential problems come from a tool in the chain that does not read or re-sign C2PA, or from a manifest nobody read before upload. The table maps what you see to the likely cause.

What you see Likely cause What to do
No credential on the delivered file An encoder, rewrap or ad server made a new file without signing Sign at the final encode, use an external manifest, or note the gap in the log
Credential present, signature invalid The file changed after signing, for example in a metadata tool Re-export from the signed master and avoid post-export tools
An unexpected AI ingredient A generated placeholder or concept frame stayed in the timeline Remove it, or disclose it if it is in the final picture
A YouTube AI label you did not expect The upload carried C2PA metadata, and that label cannot be adjusted Read the manifest before upload; correct a wrong manifest at its source
No “Captured with a camera” label The file was edited, or the camera lacks C2PA 2.1 support Expected for finished ads; no action needed
Validation fails years later The signing certificate expired and there was no time-stamp Ask your tool vendor about trusted time-stamps
Names, accounts or frames exposed Identity preferences or ingredient thumbnails were included Redact or clean-slate sign, with counsel’s agreement

One pattern deserves a warning. If a manifest overstates AI use, the fix is to correct the manifest at the tool that wrote it, not to strip all metadata before upload. Stripping removes the evidence that a step was minor, and it does not change what the rules require.

What to do next

Treat Content Credentials as one layer of the record, not the record itself. Before your next spot goes out:

  • Decide, per campaign, whether masters will be signed, by which tool, and with whose identity.
  • Turn on credential checks at import and write the AI steps into the production plan before the edit starts.
  • Validate every version after export and after the ad server, and upload one unlisted test per platform.
  • Keep a provenance log with shot, frames, tool category, reviewer and platform test results, and hand it over with the rights sheet.
  • Run the AI disclosure checklist for each version and let counsel make the labeling call.

Finishing is where most of this happens: the edit, the encodes and the QC pass described on post-production. If you want a production plan that includes a disclosure and provenance plan for your own spot, send a brief. More on AI steps in production is collected in the AI in production hub.