One Deepfake Ad Can Need Four Different Release Plans

Take one release ticket: a 30-second vertical campaign video in which software changes a candidate’s mouth and voice to deliver words the candidate never said. The edit is obvious parody to its creators, but the file contains no disclosure inside the picture or audio. A distribution team plans to post it organically, buy social ads, send it to supporters, and cut an audio version for automated calls.

That ticket cannot be cleared with a single field marked “AI labeled.” Washington specifies disclosure language for synthetic media in electioneering communications close to an election. Texas focuses on deceptive video published within a shorter window and ties liability to intent. Florida requires a generative-AI disclaimer on covered political advertising without using the same synthetic-media test.

Michigan reaches manipulated images, audio, and video and prescribes its own notice structure.

The operational unit is therefore not the campaign. It is the asset-channel-jurisdiction combination, with a record of which rule was checked and which version of the file was released.

The same file changes legal character at the state line

Washington’s synthetic-media law covers certain electioneering communications distributed within 90 days of an election. Its required notice says: “This image, video, or audio has been manipulated or generated by artificial intelligence.” The law also addresses how long a visual disclosure must appear and how an audio disclosure must be delivered. That turns a policy sentence into an export requirement: the editor needs a persistent, readable treatment for video and a spoken treatment for audio.

Michigan also uses a 90-day election window, but its law is framed around materially deceptive media and a disclosure stating that the content was manipulated by technical means and depicts speech or conduct that did not occur. The distinction matters. A generic badge reading “Made with AI” may satisfy an internal transparency policy while missing the state’s prescribed substance, which tells viewers what the manipulation did.

Texas takes another route. Its election law targets a “deep fake video” created with artificial intelligence, published or distributed within 30 days of an election, with intent to injure a candidate or influence an election. The statute is narrower in format and timing than rules that expressly include still images and audio. It also creates a criminal offense rather than relying only on a candidate’s request for correction or a private lawsuit.

Florida’s 2024 rule applies a different trigger to covered political advertisements and other election communications created in whole or in part with generative AI. The required statement is: “Created in whole or in part with the use of generative artificial intelligence (AI).” That language describes production, while Washington’s notice describes manipulation or generation. They are not interchangeable merely because both mention AI.

For the release ticket, four exports may be cheaper than one supposedly universal label. A universal treatment can still fail if it omits required words, appears for too little time, becomes unreadable in a vertical crop, or lacks the spoken notice required when the picture disappears. Maintaining derivatives adds editing and quality-assurance work, but publishing one master file creates a harder failure to detect: the team may possess a disclosure and still lack the applicable disclosure.

Timing belongs in the publishing system

Election windows should be machine-readable fields, not dates remembered by a media buyer. The Texas example uses a 30-day window; Washington and Michigan use 90-day periods for relevant provisions. Florida’s advertising disclaimer does not fit that same near-election pattern. Local contests, primaries, conventions, special elections, and runoffs can also produce different reference dates.

A practical release system stores the jurisdiction, contest, relevant election date, planned first publication, and scheduled final impression beside the asset. The last field matters because an ad approved outside a statutory window may keep running into it. Evergreen uploads create the same problem: an old post can remain accessible while a campaign pays to redistribute it closer to voting.

The ticket should trigger a fresh review when targeting expands or the schedule moves. Copying approval from a national ad into a state-specific buy is weak control because the operative facts have changed, even if the pixels have not.

Do not assume that passage settles enforcement. California enacted AB 2839 in 2024 to address materially deceptive election content across a broader period, but a federal court preliminarily blocked much of the measure during First Amendment litigation. California also enacted separate duties for large online platforms under AB 2655, which has faced its own litigation. A compliance table must distinguish enacted text, effective provisions, injunctions, pending challenges, and proposals; putting every bill into one column labeled “law” hides the status distribution teams need most.

Covered media is a technical specification

“Deepfake” is an unstable production label. Statutes instead define covered outputs through combinations of media type, deceptive realism, depicted conduct, generative method, intent, consent, and electoral purpose. A face swap, cloned voice, synthetic photograph, edited subtitle, or conventional composite may land differently under those tests.

The release ticket needs to record what changed. For the hypothetical video, the useful entry is not “AI used.” It is “mouth movement regenerated; voice cloned; candidate depicted saying words not spoken; visual and audio outputs distributed separately.” That description lets reviewers test a video-only law against the master file and an audio-inclusive law against the robocall cut without rerunning the creative investigation.

It also exposes channel-specific rules outside deepfake statutes. The Federal Communications Commission has said AI-generated voices count as “artificial or prerecorded voice” under the Telephone Consumer Protection Act, so an audio clone used in a robocall enters an existing consent and identification regime even where a state election-deepfake provision does not cover audio. Federal campaign rules, state advertising rules, defamation law, rights of publicity, and platform policies may create additional constraints. The state deepfake map is one layer, not a clearance certificate.

Platforms face a second classification task. Some state provisions focus on the person who creates or distributes the communication; others establish procedures, protections, or duties for publishers and online services. A platform may also impose a broader synthetic-media label through its own policy. That label can reduce user confusion, yet it does not automatically reproduce statutory wording or cure a missing disclosure embedded in the creative.

Remedies determine the escalation path

A rule backed by a criminal penalty calls for a different release gate from one that mainly lets a depicted candidate seek an injunction, which is a court order to stop distribution. Civil damages, statutory penalties, takedown procedures, and correction opportunities also produce different incident plans.

The distinction changes what happens after publication. If a platform crop removes the disclosure, the fallback may be to pause delivery and replace the render. If the asset falls inside a criminal prohibition tied to intent, adding a late label may not resolve the underlying issue. Where a candidate can seek an injunction, delay has its own cost because a court dispute during a short election window can consume the useful life of the communication.

Receipts matter here. Keep the source file, model or editing-tool record, human approval, applicable rule version, disclosure text, rendered previews, targeting settings, publication timestamps, and any later changes. Provenance, meaning evidence about where media came from and how it was altered, helps reconstruct the decision; it does not prove that a disclosure was legible after a platform transcoded the upload.

That is why the release ticket should include screenshots or captured playback from the destination, not only the exported master. Automated visual checks can flag a missing text layer, while speech-to-text can look for an audio notice, but compression, music, unusual fonts, and incorrect transcription create false passes and false alarms. Human playback remains the fallback for high-risk election assets.

A distribution checklist built around the asset

Start with the depiction record. Identify the person shown or heard, the speech or conduct that did not occur, the tools used, and whether the result could reasonably appear authentic. Preserve the unaltered source and project files.

Next, split the plan by channel and format. Organic video, paid video, still-image derivatives, audio calls, television, email, and supporter toolkits should not share approval merely because they came from one creative concept. Each derivative gets its own asset identifier.

Attach geography and timing to every placement. Record the relevant contest and election window, then configure an expiration or review event before a scheduled campaign crosses into a stricter period. Geofencing reduces exposure but is imperfect because users travel, share files, and receive content outside the intended target.

Render the exact applicable notice. Check wording, position, duration, contrast, font size, and spoken delivery against the operative text, then inspect the published version on a phone and desktop. Save evidence of that inspection with the ticket.

Route by remedy and uncertainty. A criminal provision, an unsettled injunction, or ambiguous coverage deserves a higher approval tier than a correctable internal-policy mismatch. This is an operational framework, not legal advice; campaigns and platforms still need qualified review for the jurisdictions and facts they face.

Finally, monitor changes. State legislatures amend election codes, courts pause provisions, agencies interpret adjacent communications rules, and platforms revise ad policies. The owner of the release ticket should receive those updates and be able to identify every live asset affected, rather than emailing the entire organization and hoping the right buyer notices.

Questions people ask

Is one national

AI disclosure enough for every political ad?

No general federal label replaces state-specific requirements. A national disclosure may exceed some platform policies but still miss prescribed state wording, audio delivery, duration, or placement. Teams can use a common transparency treatment as a baseline, then create jurisdiction and channel variants where operative rules demand something different.

Does a parody label keep a synthetic campaign video outside deepfake laws?

Not automatically. Some laws contain satire or parody protections, but coverage can depend on whether a reasonable viewer would recognize the treatment, what the disclosure says, and how realistically the person is depicted. Record the creative rationale, then test the actual render and distribution context rather than relying on the file name or the creator’s intent.

Do these rules apply only to video?

No. Texas illustrates a video-focused statute, while Washington and Michigan expressly address combinations of image, audio, and video. An audio derivative can also trigger robocall rules. Inventory every output produced from the source project because removing the picture does not remove the regulatory issue.

What should a platform save after approving an AI political ad?

Save the uploaded and delivered versions, disclosure review, targeting and timing settings, policy decision, applicable rule status, publication timestamps, and later edits. Destination playback is important: it shows whether cropping, compression, or transcription changed the notice after upload, which the pristine master file cannot establish.