A Deepfake Label on One Platform May Fail Everywhere Else

Start with a 20-second campaign video. It shows a recognizable candidate appearing to say words the candidate never spoke, using synthetic audio and photorealistic generated video. The campaign wants to upload the same file to Meta, Google and YouTube, send it to a television station, and let supporters repost it.

One editor could add “AI-generated” in the lower corner and assume the job is finished. It is not. Meta may impose its own advertiser disclosure, a state may prescribe different words or timing, a broadcaster may eventually face a federal rule that does not cover the social upload, and a Content Credentials badge may document the file’s origin without satisfying any visible-label requirement.

The useful artifact is not one universal sticker. It is a disclosure record attached to the 20-second video, showing what was generated, where the file will run, which rules were checked and which version each outlet received.

Start with the publisher, not the pixels

The first branch in the decision tree is distribution. A paid placement bought through an advertising system is governed differently from an organic post, while a television spot can sit under election law, station practices and communications regulation that never reaches a supporter’s repost.

Record every planned route before reviewing the creative: paid social placement, unpaid campaign account, creator repost, search or video ad, broadcast spot, text message, email and campaign website. Do not collapse “online” into one channel. TikTok prohibits paid political advertising, for example, so adding a synthetic-media label does not make the 20-second video eligible to run there.

Next classify the message. An ad asking voters to support a candidate or ballot measure may trigger election-specific rules. A commercial ad that uses a synthetic celebrity endorsement raises deception, publicity and endorsement issues, but the United States does not have a general federal requirement that every AI-assisted commercial carry an “AI-generated” label. A platform can still demand one under its contract with the advertiser.

Finally, identify the depicted person, the election and the states where the communication will be distributed. State election statutes often use defined windows before an election, distinguish candidates from other speakers, and require more than the loose industry meaning of “deepfake.” Some cover realistic synthetic media that falsely depicts conduct or speech; others focus on knowing distribution, intent to influence an election or paid political advertising.

Platform labels are conditions of access

Meta requires advertisers to disclose certain social issue, election or political ads when they contain a photorealistic image or video, or realistic-sounding audio, that was digitally created or altered in specified ways. Meta can place information on the ad and may reject an advertiser that fails to make the required declaration. That is platform enforcement: the practical penalty can be rejection, restricted advertising access or account action.

Google likewise requires election advertisers to disclose synthetic or digitally altered content that inauthentically depicts real or realistic-looking people or events. Depending on the ad format, Google may generate a disclosure from information supplied by the advertiser, while other formats require the advertiser to place a clear and conspicuous disclosure in the creative.

YouTube also has a creator disclosure flow for realistic altered or synthetic content. That setting matters for an organic upload, but checking it does not prove that an election advertiser completed Google’s separate ad workflow, and neither action establishes compliance with a state statute prescribing its own wording.

Return to the campaign video. The production file needs one row for Meta’s advertiser declaration, another for Google’s election-ad requirement and a separate entry for the organic YouTube upload. A screenshot of the Meta label is evidence that Meta displayed a label on that placement. It is not evidence that the television copy or a supporter’s downloaded repost carried the disclosure required elsewhere.

State election laws can prescribe the words

State rules are the branch most likely to make a generic “AI-generated” caption inadequate. Washington’s election law, for example, covers certain synthetic-media electioneering communications distributed within a defined pre-election period and supplies disclosure language tied to the format: “This image has been manipulated,” “This audio has been manipulated,” or “This video has been manipulated.” The statute also creates a route for affected candidates to seek relief.

Michigan’s rules for certain political advertisements use another formulation: “This ___________ was generated in whole or substantially by artificial intelligence.” The blank identifies the relevant medium. The difference is substantive. “Manipulated” and “generated in whole or substantially” describe different statutory tests, and a platform’s small interface label may not match either requirement’s placement, duration or readability rules.

Other states take different approaches. Texas restricts certain deepfake videos published with intent to injure a candidate or influence an election close to Election Day, rather than establishing a universal label for every synthetic political communication. California has enacted election-related synthetic-media measures covering specified periods and conduct, but parts of that state’s regime have faced First Amendment litigation. A compliance table should therefore carry a status field such as enacted, effective, challenged, enjoined or proposed, with a source and review date.

Geographic targeting does not eliminate the problem. An ad platform can restrict a paid audience by state, but campaign accounts, embeds, downloads and supporter reposts can move the same file outside that boundary. The fallback is either a disclosure that satisfies the strictest applicable rule without becoming misleading, or separate exports for separate distribution routes. That choice costs review time and version-control work, but it is cheaper than guessing from the label visible in one ad preview.

Federal proposals are not current mandates

There is no general federal rule requiring every deepfake or AI-generated political ad to display one standard label. Existing Federal Election Commission rules address specific conduct, including fraudulent misrepresentation of campaign authority, but they do not create a broad synthetic-media labeling code for all campaign communications.

The Federal Communications Commission has proposed disclosure requirements for AI-generated content in political advertisements carried by regulated broadcasters and related providers. A proposal is not an operative requirement merely because an agency published it or opened a rulemaking. Teams should record the proposal, monitor its status and avoid representing it as an enforced rule until the agency adopts an effective final requirement that covers the outlet and communication at issue.

The FCC’s treatment of AI-generated voices in robocalls illustrates another classification problem. The agency has said those voices fall within restrictions on artificial or prerecorded voice calls under the Telephone Consumer Protection Act. That concerns consent and calling practices; it does not create a universal visual label for the campaign’s 20-second video.

Provenance is evidence, not a legal safe harbor

Content Credentials, based on the C2PA technical standard, can attach cryptographically signed provenance data recording information about a file’s creation and edits. A compatible service may display that history through an icon or inspection panel. This can help a publisher show that a synthetic element was declared and that the signed file has not changed since the credential was applied.

It is a voluntary technical signal unless a contract or rule specifically requires it. It can also disappear when someone takes a screenshot, strips metadata or passes the media through a service that does not preserve the credential. A visible disclosure may survive those operations if it was rendered into the pixels or audio, although cropping and clipping can still remove it.

For the campaign video, preserve the signed master and export platform-specific copies from it. Store the script, consent records, generator or editing method, human approver, disclosure text, target states, publication destinations and screenshots of the live placements. Hashing the final file, which means calculating a repeatable digital fingerprint, helps connect the reviewed copy to the published asset without claiming that the hash proves the underlying depiction is truthful.

The record should also separate the authority behind each action. Mark Meta and Google entries as platform policy, Washington or Michigan entries as state law, an unresolved FCC item as proposed federal regulation, and Content Credentials as voluntary provenance. Add review dates because election windows, litigation and platform interfaces change.

That worksheet is the anchor when the 20-second video branches into six copies. The editor can see which words must be rendered into the file, the media buyer can complete platform declarations, and the reviewer can identify the remaining gap rather than treating a single “AI-generated” badge as universal compliance. This is an operational framework, not legal advice; campaigns and publishers still need current jurisdiction-specific review.

Questions people ask

Does a

Meta or YouTube AI label satisfy state election law?

Not by itself. A platform label shows that the platform applied its policy, while a state statute may require different wording, size, duration, timing or placement. Keep evidence of the platform label, then compare the actual state-law requirement with the version distributed in that state.

Do all

AI-assisted political ads need a deepfake disclosure?

No. Coverage often depends on whether the media realistically depicts a person doing or saying something that did not occur, whether it is paid, when it runs and which jurisdiction receives it. Routine editing or assistive use may fall outside a rule, although a platform can use a broader standard.

Are

Content Credentials enough to disclose synthetic media?

Content Credentials can document provenance and edits, but they are not a universal substitute for a visible or audible disclaimer. They may be stripped during reposting or transcoding, and viewers need compatible software to inspect them. Preserve the credential as evidence while applying any separately required disclosure.

What should a synthetic-media compliance record contain?

For each file, record the depicted person, synthetic elements, distribution channel, election and jurisdictions, applicable rule, required wording, placement, reviewer and approval date. Preserve the final export, its hash, platform declarations and screenshots of live labels so the team can show which requirement it addressed.