Political Ad AI Labels Disappear When Campaigns Repost Them

Take a common campaign handoff: an editor exports a 30-second, 1920-by-1080 video with an AI disclosure in the lower-right corner. The media team approves that file, crops it into a vertical version for a mobile feed, then a social manager downloads the posted video and uploads it elsewhere.

The underlying political message has not changed. The evidence attached to it has.

A center crop may cut off the visible disclosure. The download usually separates the video from the first platform’s ad record, including any synthetic-media declaration selected during submission. Embedded provenance may survive an ordinary transcode, disappear during export, or become inaccessible after a screenshot or screen recording, depending on the software and platform.

That 30-second spot is the useful unit of analysis because “labeling” describes several mechanisms that fail differently. A campaign can satisfy a platform form without placing words in the video, burn a notice into the pixels without preserving its editing history, or attach cryptographic provenance that most voters will never open. Treating those methods as interchangeable produces an approval record that looks complete until the asset leaves its original placement.

Platform rules attach to a particular upload

Meta and Google require disclosures for certain political ads containing realistic synthetic or materially altered media, but their policies do not create a portable label that follows a file across the internet.

Meta’s published policy says advertisers must disclose when a social issue, electoral or political ad contains “a photorealistic image or video, or realistic sounding audio, that was digitally created or altered” in specified ways, such as depicting a real person saying something they did not say. Meta can add information to the ad based on that declaration. The policy also distinguishes qualifying manipulation from edits such as cropping, resizing and color correction that are generally inconsequential on their own.

That distinction can be misunderstood. Cropping does not newly turn an ordinary video into a deepfake, but cropping an already disclosable synthetic video does not remove the reason for disclosure. It may remove the words that told the voter.

Google’s election advertising policy likewise requires covered advertisers to “prominently disclose” synthetic or digitally altered content that inauthentically depicts realistic people or events. Its stated placement test matters more than a design team’s preferred font: the disclosure must be clear and conspicuous and appear where users are likely to notice it. Google may generate a disclosure for some ad formats, while other placements require the advertiser to supply one.

These are platform requirements, enforced through ad review, account controls and possible removal or suspension. They are not universal technical standards. A declaration made in Meta Ads Manager does not populate a Google Ads submission, and neither declaration becomes part of a downloaded MP4 unless the platform or advertiser also renders a notice into the media.

Government requirements remain fragmented by jurisdiction and medium. The Federal Communications Commission’s 2024 proceeding proposed disclosure rules for AI-generated content in certain radio and television political advertising; a proposal is not the same thing as an operative cross-platform requirement. State rules vary, while the Federal Election Commission has not created a general technical labeling system that travels with political media. Campaigns need current counsel for the jurisdictions and channels they use.

This analysis is not legal advice.

The visible notice survives only if the pixels survive

Return to the 1920-by-1080 master. If the editor burns “AI-generated audio” into the image, meaning the text becomes part of every video frame, ordinary compression will usually preserve it. Voters can see the notice without opening a menu or trusting the platform to interpret metadata.

Visibility has a cost. A lower-corner label competes with captions, platform controls and the candidate disclaimer already present in many ads. Shrinking a horizontal spot into a feed preview can make technically present text unreadable. A vertical center crop can remove it altogether, while an automated “smart crop” may reposition the subject and leave the disclosure outside the selected frame.

The fallback is mundane: render the disclosure separately in every approved aspect ratio and inspect the final file at the size and duration a voter will encounter. Keeping the text within a conservative safe area reduces accidental clipping, though no single safe area fits every platform interface. A notice shown for only the opening second can also disappear when a repost starts later or a publisher trims the lead-in.

Visible text is weak evidence of origin. Anyone can add it, remove it or cover it. It remains the strongest of the three methods for immediate notice because it asks nothing from the viewer beyond looking at the ad.

For the 30-second spot, the compliance artifact should therefore be the final vertical export with its own disclosure, not an instruction that points back to the approved horizontal master. The same rule applies to a square cut, a subtitled version and a shortened excerpt. Each is a distinct public communication even when the campaign’s asset system groups them under one creative concept.

Platform metadata is useful but stays on the platform

Platform metadata records facts associated with an upload or ad object rather than drawing them into the video. It can include the advertiser’s synthetic-content declaration, account identity, ad identifier, review state and delivery history. This record gives a platform something structured to enforce and gives investigators a path back to the purchaser.

It works well inside the boundary where it was created. A platform can place its own notice next to the ad, retain the declaration in an advertising library, or use the field during review. The campaign avoids permanently covering part of the image, and the platform can change its presentation without requiring a new video export.

Reposting breaks that connection. Downloading the 30-second spot generally produces a media file, not the original ad object and its account-level fields. Uploading that file through a different account creates a new object, where the disclosure must be made again if the new placement requires it. Copying a caption is not equivalent to copying the submission record.

Compliance teams should capture the platform’s ad preview after the declaration has been applied, followed by the live presentation when available. Preserve the ad ID, destination URL, submitting account, disclosure selection and review receipt alongside the exact file delivered. A screenshot alone cannot prove which binary file was uploaded, while a file alone cannot prove which platform controls were selected.

Embedded provenance records history, not voter attention

Embedded provenance offers a third layer. The Coalition for Content Provenance and Authenticity, or C2PA, publishes a standard for attaching a cryptographically signed manifest to media. The manifest can identify the signing organization, record supported edits and link a derivative to source ingredients without claiming that the depicted event is true.

That last limit matters. A valid signature can show that a known tool or organization made assertions about a file and that the signed data has not changed unnoticed. It cannot establish the accuracy of a campaign’s political claim, and the absence of credentials does not prove manipulation.

Provenance can survive editing when every tool in the chain reads the prior manifest and writes a new one that preserves the relationship. Unsupported exports may discard it. Social platforms routinely resize, recompress or repackage uploads, and a screen recording captures displayed pixels rather than the original manifest. Some provenance systems can recover credentials through matching services or durable bindings, but recovery depends on participating infrastructure and a verifier that knows to look.

Voters also need an interface. A C2PA icon or “Content Credentials” control can open useful details where supported; an invisible manifest without a surfaced indicator provides no immediate disclosure. Provenance therefore strengthens the receipt trail more than the first glance.

For the campaign spot, retain the signed master, its manifest and every derived export before upload. Compute a cryptographic hash, a fixed fingerprint that changes when the file changes, for each delivered asset. The hash helps match an archive copy to a later file, although any platform transcode will produce a different hash and should be recorded as a separate artifact if obtainable.

Archive the handoff, not just the creative

The practical control is a release bundle for each placement. It should connect the final media file to the visible disclosure, the platform submission and any provenance record, rather than assuming one layer can reconstruct the others later.

Start with the exact export sent to the platform and a hash of that file. Keep the source project or signed master under controlled access, then record the crop, duration and other transformations used to make the placement variant. Save evidence of the platform declaration with its ad ID and review outcome. Finally, capture how the ad appeared to a viewer, since a submission screenshot cannot reveal that controls covered the notice or that the feed rendered it too small to read.

This adds storage and review time, especially when one concept produces many placements. It is still cheaper than asking an editor months later to recreate which lower-right label appeared on which crop. The approval gate should reject a variant when its notice is absent or unreadable, even if the horizontal master passed review and the platform form remains checked.

The 30-second video then has four connected records: the approved source, the exported variant, the platform transaction and the observed presentation. If the vertical repost loses its label, the archive shows whether the failure occurred during cropping, submission or platform rendering. That is an auditable answer rather than a general assurance that the campaign “used disclosures.”

Questions people ask

Does a platform’s

AI disclosure stay with a downloaded political ad?

Usually not. The declaration belongs to the platform’s ad object or upload record, while a download normally contains only the media. A campaign reposting the file should treat the new upload as a separate submission and preserve evidence of the disclosure selected there.

Is a visible label enough to document compliance?

A visible label gives voters the most direct notice, but it does not record who added it or which submission settings were used. Archive the labeled final file with its hash, the platform ad ID, the disclosure selection and a capture of the rendered ad.

Do

C2PA Content Credentials prove a political video is true?

No. Content Credentials can authenticate signed assertions about origin and editing history, provided the credentials remain available and verify correctly. They do not decide whether a political statement is accurate, and missing credentials alone do not establish that media is deceptive.

Where should an AI disclosure appear on a resized video?

Place it inside the final composition where platform controls, captions and cropping will not obscure it, then inspect the rendered placement on a phone-sized screen. Do not rely on the location approved for a horizontal master; create and review a disclosure for each vertical, square or shortened export.