Campaign Deepfake Labels Change at Every State Line

Start with one review ticket: a 20-second campaign video shows an opposing candidate at a fictional event, combines altered footage with a cloned voice, and is scheduled for connected television and social platforms. The creative team calls it parody. The media buyer wants one national file. The compliance reviewer cannot approve either description without more information.

That ticket exposes the patchwork. Some state laws focus on media realistic enough to make a reasonable person believe an event occurred. Others reach political advertising created partly with generative AI, even when the manipulation is disclosed by the surrounding joke. Several operate only during a defined pre-election window.

A disclosure can provide a defense or safe harbor in one jurisdiction, while another state may prohibit distribution when intent and timing conditions are met regardless of the label.

The result is an engineering constraint. Before rendering the final file, a campaign or publisher must identify what was generated, who appears to speak or act, where the audience is located, when the communication will run, and whether the distribution method triggers separate rules. This article describes a conservative operational workflow based on public statutes, agency materials, and reporting. It is not legal advice, and current statutory text and court orders require review before publication.

The same file can trigger different tests

California’s established candidate-deepfake provision illustrates a disclosure-centered rule with a timing threshold. California Elections Code Section 20010 addresses materially deceptive audio or visual media distributed within 60 days of an election when specified knowledge, intent, and candidate conditions are present. Its prescribed disclosure says: “This _____ has been manipulated.” The law also addresses how long a visual disclosure remains visible and when an audio disclosure must be spoken, so a caption placed only in a social post is not equivalent to embedding the statement in the media.

California has enacted additional provisions covering deceptive election content, and those newer rules have faced First Amendment litigation. That distinction matters operationally: enacted text, an injunction affecting enforcement, and a bill still moving through a legislature are three different statuses. A compliance table should record each one rather than reducing them to a green or red cell labeled “legal.”

Florida takes a broader labeling approach for covered political advertising made with generative AI. Its prescribed disclaimer reads: “Created in whole or in part with the use of generative artificial intelligence (AI).” That wording describes the production method, while California’s quoted language describes manipulation. Substituting a campaign’s preferred phrase, such as “AI-assisted content,” may sound clearer to a creative director but does not reproduce either statutory statement.

Wisconsin uses another formulation for covered political communications containing synthetic media: “This [audio/video/image] has been created or intentionally manipulated by digital technology.” Its rule also distinguishes visual placement from spoken audio treatment. Washington’s synthetic-media law centers on deceptive media depicting a candidate, distribution near an election, and the absence of a specified disclosure, while providing a civil enforcement route and addressing some publisher or distributor circumstances.

Minnesota demonstrates why a universal label is an unsafe assumption. Its election deepfake provisions focus on dissemination near an election, the realism of the false depiction, knowledge, and intent to injure a candidate or influence an election. A disclosure should not be treated as a cure unless the applicable law makes it one. For the 20-second review ticket, adding “AI-generated” to the first frame may improve transparency without resolving whether distribution itself creates exposure under a prohibition-oriented statute.

Audio needs its own review path

The cloned voice in the ticket cannot ride through the video checklist as an incidental component. A synthetic voice can independently satisfy a law’s definition of deceptive audio, and a silent visual label will disappear when someone hears the file through a phone call, podcast insertion, radio spot, or audio-only repost.

The reviewer should export a transcript and mark every synthetic or materially altered passage. The production record should identify whether the system generated the words and voice, transformed a real recording, or merely cleaned noise. Routine editing does not automatically become a deepfake, but the legally relevant line often turns on whether the result falsely depicts speech or conduct and appears authentic to a reasonable listener.

Channel rules then sit on top of state election law. The Federal Communications Commission has said that AI-generated voices count as an “artificial or prerecorded voice” under the Telephone Consumer Protection Act. That ruling does not create a general deepfake-label safe harbor; it places AI voice calls inside an existing telephone-law framework involving consent and calling restrictions. Broadcasters, streaming services, and ad platforms may impose further submission or disclosure requirements.

For the ticket, the fallback is two deliverables rather than one. The video version carries an on-file visual statement for its full required duration where applicable, while the audio master includes a clearly spoken disclosure positioned to satisfy the strictest confirmed rule for its destinations. The campaign still needs jurisdiction-specific review because exact wording and repetition requirements can conflict.

Timing belongs in the asset record

Election proximity is not metadata that the buyer should add at launch. It can change the legal treatment of the creative.

The review ticket needs the relevant election, the candidate’s ballot status, the first and last distribution dates, and any primary, special, or general election that falls inside the campaign window. California’s established provision uses a 60-day period for covered candidate media. Other states use different windows, including periods measured around 90 days, while some disclosure rules for political advertisements are not limited in the same way.

A scheduler should recalculate that field whenever a flight is extended. An asset approved for an earlier awareness campaign can cross into a regulated window without anyone editing the file, which means the compliance change comes from the calendar rather than the model output. The practical fallback is an automated hold before the threshold, followed by a human reapproval tied to the new run dates.

Keep both timestamps. “Approved in June” does not establish what version ran later, and “posted before the window” does not answer whether a platform continued delivering, recommending, or republishing it inside the window.

Distribution changes who must stop the file

The social upload and connected-television placement in the ticket should be separate rows. State statutes vary in their treatment of a person who produces media, a campaign that pays to distribute it, a publisher that carries it, and an intermediary that only hosts or transmits it. Exemptions for bona fide news coverage, satire, or a broadcaster that identifies a manipulation are also jurisdiction-specific and fact-dependent.

Platform policy adds a private enforcement layer. Meta requires advertisers to disclose certain digitally created or altered content in political or social-issue ads. Google also requires disclosure for election advertising containing synthetic or digitally altered content that inauthentically depicts real or realistic people or events. TikTok prohibits paid political advertising rather than offering campaigns a labeling route.

A state-compliant file can therefore be rejected by a platform, while platform approval says nothing conclusive about state law.

Publishers need the same asset record, especially when repackaging campaign material in reporting. A news segment that explicitly examines a fake clip is different from an uncaptioned embed that a reader could mistake for source footage, but an exemption should be mapped to the publisher’s actual presentation. Preserve the headline, player treatment, surrounding disclosure, publication time, and received source file.

A conservative labeling workflow

The review ticket should begin before editing ends. Assign the asset an identifier, retain the source recordings, and record the generation or alteration tools used. Save a transcript for audio and a shot-level note for video. A cryptographic hash, a short fingerprint calculated from a file, can later show whether the reviewed export matches the distributed one.

Next, classify the depiction. Record whether it shows a real candidate, invents speech or conduct, changes only cosmetic details, or presents an obviously fictional scene. Do not let “parody” replace those factual fields. Humor may matter to legal analysis, but it does not tell an ad server which file to deliver or prove that a viewer would recognize the joke.

Build the routing matrix around destination state, election date, medium, payer, and publisher. Each applicable rule should include its status, triggering conditions, exact disclaimer, placement requirement, enforcement mechanism, and any confirmed exemption. Proposed legislation belongs in a watch column. Enjoined text belongs in a separate status field with the relevant court record.

Then render variants. Put the required statement inside the image or video rather than relying on post copy, and add a spoken version to audio when the rule calls for one. Preserve enough contrast, size, volume, and duration for the disclosure to be perceived in the delivered format. Re-encoding can crop captions, connected-television interfaces can cover lower-screen text, and platform audio normalization can make a rushed disclosure harder to hear.

Finally, require a human release decision and preserve the receipt: reviewed file hash, destinations, run dates, statutory version checked, platform declaration, approver, and withdrawal contact. If the team cannot establish those fields, the conservative fallback for the 20-second ticket is to pause distribution or replace the fictional depiction with authenticated footage and ordinary narration.

Questions people ask

Does one

“AI-generated” label satisfy every state law?

No. States prescribe different wording and treatment for audio or visual media, and some rules focus on deceptive depiction rather than the use of AI itself. A generic label may also fail placement, duration, or audibility requirements. Use destination-specific variants when the confirmed requirements do not share a workable statement.

Can a campaign label a deepfake and run it anywhere?

A label may supply a defense or satisfy a disclosure duty under some statutes, but it does not erase every prohibition tied to intent, realism, candidate harm, or election timing. Platform rules can also block the placement. Review the distribution decision separately from the disclosure design.

Do publishers need to review campaign deepfakes they did not create?

Yes, when they distribute or repackage the material. Some laws contain protections for news reporting, broadcasters, or intermediaries, but the conditions vary. Publishers should retain the source file and document the contextual disclosure shown to readers instead of assuming that editorial purpose creates an automatic exemption.

What should a team save after approving synthetic campaign media?

Keep the exact distributed file and its hash, source materials, transcript, alteration notes, destinations, run dates, required disclosure, platform submission, and named human approval. Those records cannot prove that every legal judgment was correct, but they show which asset was reviewed and stop an outdated approval from silently following a revised file.