A Deepfake Label That Works in One State May Fail Next Door

Start with one publishing ticket: a campaign supplies a 30-second vertical video in which a candidate appears to say words the candidate never spoke, and a media team plans to place it on social platforms, connected television and a campaign website. The clip includes a small “AI-generated” watermark in the lower-right corner. Distribution will cross state lines.

That watermark may be useful provenance, meaning information about where media came from and how it was altered. It is not a compliance strategy. State laws can prescribe different words, require a notice to remain visible throughout a video, cover only communications near an election, or attach obligations because a campaign paid to distribute the clip. Some laws target deceptive synthetic media even when no paid advertisement exists.

The practical unit of work is therefore the publishing ticket, not the file. Before anyone exports the video, the ticket needs enough information to route that file through every potentially applicable rule.

The file does not reveal its legal category

The first field on the ticket should describe what the audience will receive: video with sound, audio alone, a still image, or several versions derived from the same source. A disclosure burned into the video frames will disappear when an editor exports the soundtrack as a radio spot. A spoken warning can be lost when a platform starts playback muted. Captions entered in a social post may not travel with the file when another account downloads and reposts it.

California’s Elections Code offers a concrete example of why the rendering matters. Its rule for certain materially deceptive candidate media specifies the statement “This video has been manipulated” for video and requires the text to appear throughout the video, in a size at least as large as the largest other text shown. For audio, the required statement is “This audio has been manipulated,” spoken clearly at the beginning and end in the same language as the recording.

A small “AI-generated” watermark does not match those words or placement conditions. Nor does a compliant video notice automatically make the audio-only export compliant. The review ticket should treat each rendition as a separate deliverable, with a screenshot or audio sample showing the disclosure after the final encode.

Washington’s synthetic-media law uses a similar medium-specific structure for covered electioneering communications, including notices that identify an image, video or audio recording as manipulated. Its visual treatment addresses readability, duration, relative type size and color contrast; audio notices must be clearly audible and placed at the beginning and end. Similar architecture does not make the rules interchangeable. Definitions, covered speakers, election periods and remedies still differ.

This is where automated checking can help, within limits. A video pipeline can detect whether required text appears in every frame, compare its approximate height with other on-screen text and flag low contrast. Speech recognition can test whether an audio notice appears near both ends. It cannot reliably decide whether a depiction is legally “materially deceptive,” whether an exception applies, or whether the intended audience places the communication inside a state’s election law.

The fallback is human review, recorded in the ticket.

Timing changes the answer

The second ticket field should record the relevant election, publication time and planned end date for distribution. “Political content” is too broad to route correctly, while “publish Tuesday” is too narrow if a platform campaign continues running for weeks.

Texas illustrates the timing problem. Its election law addresses certain deepfake videos published and distributed within the 30 days before an election, when the actor knows the video is a deepfake and acts with the specified intent to injure a candidate or influence an election. California’s candidate-media provision uses a different pre-election window. Washington connects its rule to the state’s definition of an electioneering communication, which carries its own timing and targeting conditions.

The same 30-second clip can consequently move between legal categories while its pixels remain unchanged. A media team that approved the asset when it was outside one state’s window should not let a scheduled campaign roll forward without another check. The publishing system needs a date trigger that reopens review when distribution enters a covered period.

That trigger has a real operating cost. Someone must maintain an election calendar, map special elections and primaries, and stop scheduled placements when the ticket lacks an answer. Small publishers may decide that geofencing, which limits delivery by the viewer’s estimated location, costs less than producing and verifying several state-specific versions. Geofencing is imperfect because IP-based location can be wrong, travelers move, and organic reposts ignore the original targeting settings.

A conservative production choice is to use the most demanding known disclosure across a wider campaign. That can reduce versioning, but it still does not resolve laws that define covered media differently or restrict conduct rather than offering a label-based route to publication.

Record who created the message and who bought the audience

The third field on the ticket should separate four roles: the person depicted, the organization that generated or altered the media, the publisher operating the account, and the entity paying for distribution. These roles often collapse into one campaign team, but the law does not always assume that they do.

Wisconsin’s rule for political communications containing synthetic media makes sponsorship central. Covered audio communications use the statement “Contains content generated by AI” at the beginning and end, while covered visual communications display that statement for the communication’s duration. The law sits alongside political-advertising rules, so identifying the sponsor is part of determining whether the requirement attaches.

Michigan uses another formulation for covered political advertisements: “This message was generated in whole or substantially by artificial intelligence.” Its statute addresses how the disclaimer must appear in visual and audio formats. “Substantially” matters. A campaign’s use of generative fill to remove a microphone from the background presents a different classification problem from synthesizing the candidate’s mouth movements and voice, even though both workflows used an AI tool.

For the ticket’s 30-second video, an invoice is evidence. So is the platform ad account, insertion order or agency authorization showing who financed delivery. The uploader’s name alone is not enough, particularly when a consultant posts through a campaign-controlled account or a political committee pays a publisher to amplify an existing clip.

News coverage also requires a separate path. Several state laws distinguish paid political communications from broadcasts, reports or commentary that disclose the media’s nature, but exemptions vary. A newsroom should record why it classified a segment as reporting rather than assume that placing a manipulated clip inside a news package ends the inquiry.

A workable prepublication gate

The ticket can remain one page if it forces evidence rather than broad assurances. It should identify every state intentionally targeted, the elections and distribution dates, each media rendition, the exact synthetic alteration, the depicted candidate or officeholder, and the payer. Attach the final files, disclosure transcript, frame capture and purchase record.

Next, route the ticket against a maintained rules table. Each row should hold the state, law and current status; covered media and speaker; election window; required words; placement and duration; exceptions; enforcement mechanism; and the date someone last checked the official source. Free-text notes are useful, but structured fields let the system block an export when, for example, the video notice appears only during the opening frames.

The final gate should produce one of four operational outcomes: publish the national version, create state-specific versions, exclude a jurisdiction from paid distribution, or pause for qualified review. This article is a reporting synthesis, not legal advice, and the table should link to current statutes and court orders rather than treat a news summary as controlling text.

Keep the receipt after publication. Archive the rendered file, not merely the editable project, because fonts can change and audio tracks can be replaced before export. Save platform targeting settings and campaign dates as well. If the sponsor edits the placement later, the publishing ticket should reopen instead of silently inheriting an approval given to a narrower audience.

Enacted is different from enforceable

A state legislature’s passage of a bill establishes neither active enforcement nor a settled constitutional boundary. Some synthetic-election-media laws authorize civil actions by depicted candidates, some use criminal penalties, and others place duties on sponsors or platforms. Enforcement may depend on a complaint after publication rather than advance review by an election agency.

Court orders can also suspend all or part of a law. California’s newer election-deepfake measures have faced constitutional challenges, including litigation over restrictions and compelled labels. A rules table that says only “California: disclosure required” hides the most consequential status field: which provision is in force for this communication now.

Proposals belong in a separate queue. They can justify preparing a new field or testing a disclosure renderer, but they should not be presented to editors as current law. The same separation applies to platform policies, voluntary provenance standards and enacted statutes. A platform may remove deceptive media that state law would permit, while a cryptographically signed provenance record may document editing history without displaying the words a statute requires.

Return to the 30-second video. Its lower-corner watermark can stay, but the approval rests on the ticket showing the final medium, exact notice, full display duration, election window, targeted jurisdictions and payer. If any of those fields changes, it is a new publishing decision.

Questions people ask

Does an

“AI-generated” watermark satisfy political deepfake laws?

Not necessarily. Some statutes prescribe different wording and specify size, contrast, duration or spoken placement. A watermark can support transparency while still failing a state requirement, and a label may not cure conduct covered by a prohibition. Compare the rendered asset with the current rule in every targeted jurisdiction.

Should a publisher use the strictest state label everywhere?

That can reduce the number of versions, especially when one notice remains visible throughout a video and is repeated in its audio. It does not reconcile different definitions, election windows or sponsor rules. The strictest visual treatment is therefore a production baseline, not proof that national distribution complies everywhere.

Can geofencing keep a political video outside a covered state?

It can narrow paid delivery, but location signals are imperfect and downloaded copies can travel beyond the original campaign. Keep the targeting configuration and purchase record, then distinguish controlled paid distribution from organic publication. A public website or unrestricted social post may require a different review than a state-limited ad buy.

Who should own the publishing checklist?

A production or ad-operations owner should collect the file, dates, states and payer, while a governance owner maintains the rules table and escalation path. Editors should not have to interpret statutes during export. They should see a block when evidence is missing and know who can approve, revise or stop the placement.