Political video under state deepfake disclosure laws

Rapid response is measured in hours. Something happens at 9am, and the spot that answers it is worth far more at noon than at six. Every process improvement in political video compresses that window.

Which is exactly why disclosure cannot be a step someone remembers. Synthetic media in election communications is regulated state by state, through a patchwork that has been added to and amended nearly every legislative session, and several of those statutes prescribe the size, placement and duration of the on-screen text rather than leaving the form to you. A rule that depends on a tired staffer at 11pm adding a lower third is a rule you will break, and here breaking it is not an ad rejection. It is a news story about your campaign, written by your opponent.

The fix is architectural. Disclosure lives in the export template. If the file exports, the disclosure is on it, in both layers, or the export does not happen.

What actually varies across the map

There is no single federal AI-disclosure rule for political advertising, which is why this is a state map. And the statutes are not variations on one template — they differ on the axes that determine whether your spot is covered at all.

• Who is regulated. Some reach anyone distributing covered media. Others reach only candidates, committees, or paid advertising.
• The window. Many attach only within a set number of days before an election. Outside it, the same spot may be unregulated in one state and regulated in the next.
• What counts as covered media. Definitions range from narrow — a depiction of a real, identifiable person saying or doing something they did not — to broad enough to catch ordinary digital alteration.
• The form of the disclosure. Several specify size, placement and duration of the on-screen text, sometimes tied to the size of other text in the ad. Some specify wording.
• Metadata. Whether machine-readable marking in the file is contemplated at all. Most of the pressure for provenance metadata currently comes from general AI-transparency and platform statutes rather than from election law, but the two are converging.
• Exemptions. News reporting is commonly carved out. Satire and parody are sometimes carved out only if labelled, which is a very different rule.
• Remedies. Injunctive relief, statutory damages, private rights of action, and in some places criminal exposure.

Two consequences follow, and they are the whole planning problem. You do not run one spot in one state — a rapid-response video posted to a national account is distributed everywhere, including states whose rules you did not check, and geo-targeting helps for paid placement and does nothing for organic. And the map moves: states have added and amended these laws session by session, and litigation has touched several of them, so any count you read is a snapshot. Compliance counsel maintains the map; the producer works from a dated one-page summary rather than from memory.

The only sane posture for a campaign that distributes nationally: apply the strictest rule you are subject to, everywhere. It costs a line of text and a metadata field, and it removes the question.

Two layers, and why you need both

Layer one: burned-in visible text. Not a platform toggle, not a caption in the post copy, not a description line. Pixels in the frame. Post copy is stripped the moment someone downloads and reposts, and reposting is the entire distribution model in political video.

Write it plainly. "This video contains AI-generated imagery" beats anything hedged, and writing an AI disclosure line nobody scrolls past applies directly — the goal is a line that reads as confident rather than apologetic. Where a statute prescribes wording or size, use the prescribed version and stop optimising.

Layer two: metadata. Content Credentials are the standard the industry has converged on for recording how a file was made and what happened to it afterwards; what C2PA records on design files covers the shape of that record. This is the layer the general AI-transparency statutes are pushing toward, and a campaign that already produces it is not rebuilding its export step later.

Plan around its weakness: metadata survives some upload pipelines and is stripped by others. Some platforms read Content Credentials and surface a label from them — TikTok labelling uploads with Content Credentials is one case — and others discard the block on re-encode. The metadata layer is required where it is required, and never sufficient alone.

Invisible watermarking is a third thing and substitutes for neither. Provider-specific schemes are detectable only with the provider's tooling, and a regulator asking whether your ad was labelled will not run a vendor detector.

Rebuilding the export step so both layers are produced by default rather than added by hand is the job described in machine-readable marking and your export step; the per-platform variations are in one disclosure, five destinations.

Build the template before the news breaks

The deliverable is a saved timeline, built once, in a quiet week. Everything below is pre-placed:

Track	Contents	Set once?
V1	Placeholder for the day's footage or generated plate	No
V2	Lower-third text, brand-locked	No
V3	Disclosure text, position and duration locked to the strictest applicable rule	Yes
V4	Paid-for-by / authorisation card	Yes
A1	Voiceover slot	No
A2	Bed, ducked	Yes
Captions	Burned in, style locked	Yes

The disclosure track is effectively undeletable, because deleting it is a visible act someone has to perform deliberately. That is the point. A checklist can be skipped silently; removing a track from a template cannot.

Two details matter more than they sound. Keep the timeline at the default 25 fps so mixed-source footage does not introduce a stutter you have no time to diagnose at 11pm. And there is no vendor watermark on any plan, so the disclosure line is the only overlay competing for the frame — which matters when a statute asks for it to be clear and conspicuous and your corner is already occupied by somebody's logo.

The run sheet

Once the template exists, the rapid-response cycle is roughly ninety minutes with two people.

1. Decide the claim, in one sentence. Written down and agreed before anyone opens an editor. Rapid response fails on scope, not speed.
2. Check the register. Does this spot depict a real, identifiable person doing or saying something? If yes, escalate to counsel before production. This is the category the statutes were written for and it is not a producer's decision.
3. Source the visuals. Real footage first, always. Generated material is for plates, backgrounds, data-graphic beds, maps and typographic frames.
4. Voiceover. Synthesised narration is fast and is disclosed like any other synthetic element. A synthesised voice resembling a real named person is a different act and belongs in step 2.
5. Drop into the template. V1 and A1 filled, everything else already correct.
6. Preview at 480p. Costs nothing, carries a short per-user cooldown, and confirms the disclosure sits in the safe area on a phone with a caption bar over it. The check people skip and then regret.
7. Export once. A single charge applies to the final export regardless of how many clips are on the timeline.
8. Verify before distribution. Play the export, confirm the burned-in line is legible at phone size, confirm the metadata block is present — auditing your exports with detection tools covers how that check runs.
9. Log it. Asset, date, states targeted, disclosure text used, whether generated material appeared and where. When someone asks in October about a spot from August, the log is the answer.

Keeping this as a re-renderable draft rather than a flattened file is what makes step 7 repeatable at speed — see reusable editor drafts.

What not to generate at all

Some things are not disclosure problems. They are things you do not do, and no label fixes them.

• A real, identifiable person saying or doing something they did not. The conduct the statutes exist to reach. A label does not cure it in every state and does not cure it reputationally anywhere.
• A synthesised voice of a real named person. Including your own candidate, with their permission, in an ad. The clip outlives the permission.
• Fabricated documentary evidence. A generated crowd, protest, news chyron or ballot. Presenting a synthetic depiction as a record of an event is the specific harm.
• Generated depictions of identifiable private individuals. Constituents, voters, victims. Even sympathetically. Especially sympathetically.
• Anything that could pass for a news broadcast. News exemptions protect the press. They do not protect an ad that looks like the press.

What generated material is genuinely good for here: typographic backgrounds, map and data-graphic beds, abstract plates behind quotes, animated charts, seasonal establishing footage of a location you cannot reach today, and title cards. That covers most of what a rapid-response spot needs behind a voiceover, and none of it depicts anyone. Add text overlay to video handles the typographic layer.

For advocacy organisations rather than electoral campaigns, the fundraising-side formats are in AI video for nonprofits and fundraising.

FAQ

Does the disclosure requirement apply to organic posts or only paid ads?

It depends on the state, and several reach distribution rather than paid placement specifically. Since organic political video is reshared far beyond the account that posted it and cannot be geo-fenced, treating organic as exempt is a bet against your own distribution working. Label everything.

We only used AI for the background plate. Do we still label it?

Assume yes, and check the definition in the states you care about, because several are drafted broadly enough to catch alteration well short of a fabricated person. An unnecessary label costs one line of text. A missing one costs you an opponent's press release.

Does satire get an exemption?

Sometimes, and in several states the exemption is conditional on the material being labelled — the opposite of what people assume when they hear "parody exception." Do not rely on it without checking the statute, and if you are relying on it, label anyway.

How do we handle the difference between profile-level and post-level labels?

Platform AI labels and statutory disclosure are separate obligations that happen to look similar, and satisfying one does not satisfy the other. Profile-level versus post-level AI disclosure sets out the distinction. The burned-in line is yours; the platform toggle is theirs; do both.