One Political Deepfake Ad May Need a Different State Label

Start with one 30-second campaign video. It shows an opposing candidate at a lectern, uses a cloned version of the candidate’s voice, and includes words the person never spoke. The campaign plans to buy connected-TV placements, upload shorter cuts to social platforms, and send the audio through a phone outreach vendor.

Calling that file “AI-generated” does not answer the compliance question. A deepfake, meaning synthetic or altered media that convincingly depicts a real person doing or saying something that did not occur, can meet one state’s disclosure trigger but sit outside another state’s rule because the election is too far away, the medium is excluded, or the statute requires a particular intent.

The practical unit of review is therefore not the master video. It is each combination of creative, medium, audience location, publication date, sponsor, and distribution channel.

This is an operational map based on public materials, not legal advice. State laws are changing, and litigation can separate a requirement written in a statute from one officials may currently enforce.

Build the intake record before writing the label

The first control belongs upstream of ad trafficking. When the 30-second video enters review, its record should identify every real person depicted, which frames or sounds were generated or materially altered, whether the depicted conduct occurred, and who approved the synthetic element. A checkbox marked “uses AI” is too coarse.

That record needs the source files as well. Keep the original footage, prompts or editing instructions where available, voice-model inputs, vendor output, final export, and approval history under a stable creative identifier. This is provenance, the documented origin and editing history of a file. It lets a reviewer distinguish a cloned voice from ordinary noise reduction and a fabricated scene from a conventional color correction.

Next, record the communication’s purpose and sponsor. Some state provisions attach to political advertisements or electioneering communications, while others focus on deceptive media distributed with knowledge or an intent to affect an election. An unpaid supporter’s repost, a campaign’s paid placement, and a broadcaster’s news segment can therefore receive different treatment even when the underlying pixels match.

Timing needs its own field rather than a note in the filename. Several laws use election windows, and those windows are not uniform. The same upload can fall outside a state restriction when first published, then enter the covered period while it remains available or is promoted again. A campaign should not assume that an approval issued months earlier survives a new flight date.

Return to the master ad. Its cloned voice, fabricated statement, paid campaign sponsor, and planned publication date become structured facts in the intake record. Those facts feed the state-routing decision; the editor does not choose a label from memory.

The trigger is more than whether software was used

Florida illustrates a broad, advertisement-centered trigger. Its law covers specified political advertising that was created in whole or in part with generative AI and depicts a real person performing an action that did not occur. The prescribed disclaimer says: “Created in whole or in part with the use of generative artificial intelligence (AI).” The rule also sets medium-specific presentation requirements.

Michigan uses a different construction for materially deceptive media published near an election. Its statutory disclosure language states: “This [image/video/audio] has been manipulated by technical means and depicts speech or conduct that did not occur.” The bracketed medium changes with the file, and the law includes format-specific rules for displaying or speaking the disclosure.

Those sentences are not interchangeable. Florida’s wording identifies the production method. Michigan’s wording tells the audience that the represented conduct did not happen. A national campaign that burns only the Florida sentence into its video should not infer that it has satisfied Michigan’s prescribed statement, even though both labels alert viewers to manipulation.

Other state provisions narrow or reshape the trigger through the kind of person depicted, the creator’s knowledge, an intent to deceive or influence voters, the communication format, or a period measured from an election. Some provide exceptions or distinct treatment for satire, parody, news reporting, or a publication that carries a qualifying disclosure. The details determine whether a label cures the issue, whether distribution is restricted despite a label, or whether that statute does not reach the communication at all.

This is why automated detection cannot make the final call. A classifier may estimate that a face or voice is synthetic, but it cannot reliably establish the sponsor’s intent, whether the depicted statement occurred, which election the communication concerns, or whether a statutory exception applies. Detection can open a review ticket. It should not close one.

Put the disclosure where the covered version travels

A disclosure in an ad-library record is not necessarily a disclosure on the ad. State rules may require words inside a visual communication, spoken language in audio, or a statement presented with specified prominence or repetition. Metadata can support an audit trail, but it disappears when someone screen-records a video, extracts its audio, or uploads the file elsewhere.

For the 30-second master ad, the safest production workflow creates separate rendered assets after the routing review. The connected-TV version may need a visual disclosure placed within the picture rather than in a web caption. The audio cut may require a spoken statement, which consumes paid airtime and can force the campaign to shorten the underlying message. A static thumbnail derived from the video needs its own review because a label visible later in the moving footage will not accompany the extracted image.

That cost is not theoretical. Every state-specific render adds editing time, quality assurance, trafficking complexity, and the chance that a buyer attaches the wrong file to an order. The alternative, one maximally prominent label used everywhere, is easier to administer but can take up screen space, interrupt audio, and describe content more broadly than a particular jurisdiction requires.

For many national campaigns, the broad label is still cheaper than maintaining a large set of nearly identical files. The decision depends on volume and message sensitivity, but it should be explicit: either accept a conspicuous national disclosure or fund the version control needed for narrower state treatments. Quietly accumulating ad hoc exports is the expensive option because nobody can later prove which label reached which audience.

Platforms add another ruleset. Google requires election advertisers to disclose certain synthetic or digitally altered content that inauthentically depicts real or realistic-looking people or events. Meta requires advertisers to disclose specified digitally created or altered photorealistic imagery, video, or realistic-sounding audio in political or social-issue ads. TikTok does not allow paid political advertising.

Those policies do not replace state law. A platform-applied notice may use different wording, appear outside the media, or cover a different category of alteration. The campaign’s routing record should therefore store platform policy and state requirement as separate checks, followed by a screenshot or rendered-file hash showing what was delivered.

Geography is a delivery control, not a legal conclusion

Ad systems can target states, but geofencing does not create a sealed border. Users travel, location signals can be stale, connected-TV households may be assigned through imperfect data, and a recipient can repost the creative without its original caption. Organic publication on a campaign account creates another problem because the audience is not confined to the paid targeting settings.

The campaign should record the geography it instructed each vendor to use and the evidence returned by that vendor, while recognizing that this proves a delivery setting rather than every viewer’s location. If a state-specific rule depends on distribution in the state, counsel may need to interpret what the law requires; the ad operations team should supply logs, not make that legal judgment.

The 30-second master ad now has a usable release packet: factual intake record, state matrix, effective-status field, approved wording, rendered asset identifier, destination platform, flight dates, geographic settings, and publication evidence. If the audio becomes a robocall, it receives a new route because federal telephone rules and state call requirements can apply independently of video-ad rules.

Track enacted, effective, enforceable, and proposed separately

A statute’s passage does not tell an operator whether a requirement is active. The matrix needs distinct fields for enactment, effective date, covered election window, agency guidance, pending litigation, and any order limiting enforcement. California’s synthetic-media election laws, for example, have faced constitutional challenges and court orders affecting enforcement. A row marked only “California has a law” conceals the operational fact that the written requirement and current enforcement posture may differ.

Federal proposals belong in another status category. The Federal Communications Commission has proposed AI-disclosure requirements for certain political advertisements carried by regulated broadcasters and other covered services. A proposal should inform planning, especially for contracts and asset templates, but it should not be represented in the release checklist as a final federal mandate. Existing FCC treatment of AI-generated voices in robocalls is a separate rule with a different trigger.

Assign someone to refresh this matrix before each flight, not merely once per election cycle. Public reporting can flag a new bill, injunction, or platform policy, but the reviewer should link the current official text or order and preserve the version used for approval. That receipt matters when the rule changes after the ad has run.

Questions people ask

Can one disclosure satisfy every state’s political deepfake law?

Not reliably. States can prescribe different wording, cover different media, and use different election windows or intent standards. A campaign may choose one broad national label for operational simplicity, but it still needs a state-by-state review showing that the wording and placement satisfy each applicable requirement.

Does a platform’s AI label satisfy state disclosure rules?

Not necessarily. Platform notices follow the company’s own trigger and may appear beside the post rather than inside the image, video, or audio. Keep the platform check separate from the state-law check, and preserve evidence of both the uploaded creative and the notice viewers received.

Can geofencing keep an ad outside states with stricter rules?

Geofencing can reduce intended delivery, but it cannot guarantee where every impression or repost appears. Campaigns should retain vendor settings and delivery reports, then escalate the legal significance of spillover rather than treating a state boundary selected in an ad dashboard as conclusive.

What should a campaign save after approving a synthetic ad?

Save the source and final files, alteration record, depicted-person analysis, approved disclosure, state matrix version, platform settings, flight dates, geographic instructions, approvals, and publication evidence. Tie them to one creative identifier so the campaign can reconstruct which rendered version ran in each place.