Before You Publish an AI Election Ad, Run These Checks

Start with a hypothetical 30-second spot. A campaign uses a voice-cloning tool to make its candidate read an approved script, pairs the audio with generated footage, and plans to distribute the result on television, social platforms, a campaign website, and prerecorded calls.

The creative team sees one advertisement. A compliance review sees several regulated acts: paying to place a campaign message, depicting a real person synthetically, broadcasting political material, uploading content under a platform contract, and potentially delivering an artificial voice to telephone subscribers. Each act has its own trigger.

That distinction should control the workflow. Do not ask whether “AI ads” are legal in the abstract. Build a release sheet for the specific file, then block distribution until the owner of each channel has answered the applicable disclosure, consent, and recordkeeping checks.

Freeze the file and describe what the system changed

Create the release sheet before legal or policy review. Give the spot a file identifier and record whether the system generated or altered the voice, face, body movement, setting, or sequence of events. “AI-assisted” is too imprecise: noise removal and color correction present different risks from making a candidate appear to say words never recorded.

For the 30-second spot, the sheet should state that a voice model generated the candidate’s speech from an approved script and that an image generator produced footage presented as illustrative rather than documentary. It should also identify every real person who is recognizable, including a person who appears only in source footage.

Preserve the pre-edit source, generated output, final export, script, prompts, approvals, and tool account used. Record the model and version shown by the service where available. A cryptographic hash, a short digital fingerprint that changes when a file changes, can connect the approved export to the file later delivered to publishers.

This costs staff time and storage, and some vendors expose less generation history than others. The fallback is manual capture: export available logs, take dated screenshots of settings and approvals, and write down what the tool cannot reproduce. A polished provenance dashboard is not worth buying if the team still cannot identify who approved the cloned voice.

Map every destination before reviewing the label

Make one row for each state and each delivery method. A national upload is not one row merely because it uses one file; viewers may receive it under different state laws, while television stations, websites, paid social placements, and telephone calls sit under different federal and contractual regimes.

Federal Election Commission rules already require disclaimers on many campaign communications, regardless of AI use. Under 11 CFR 110.11, a covered disclaimer must be “clear and conspicuous,” and the prescribed message changes according to who paid for the communication and whether a candidate authorized it. An authorized committee communication generally identifies the committee as the payer.

An unauthorized communication generally identifies its payer and states that it was not authorized by a candidate or candidate’s committee.

The internet trigger is narrower than many teams assume. The FEC definition of “public communication” includes a communication placed for a fee on another person’s website, digital device, application, or advertising platform; an unpaid post on the campaign’s own account may fall into a different federal disclaimer analysis. State rules and platform policies can still apply.

Keep the ordinary sponsor disclaimer separate from any synthetic-media label. One answers who paid. The other tells viewers that depicted speech, conduct, or events were generated or materially altered. Combining the text may be possible in a given format, but the release sheet should test each requirement independently so a correct “Paid for by” line does not conceal a missing AI notice.

Read the current state text, including its exceptions

State synthetic-media laws vary in their covered media, election window, intent standard, remedy, and required wording. Some target materially deceptive depictions of candidates; some permit distribution with a prescribed disclosure; others create a civil claim or address only particular formats. A candidate’s own cloned voice may therefore receive different treatment from an opponent’s fabricated concession speech.

Washington provides a useful example of why summaries are insufficient. Its synthetic-media law uses the disclosure “This (image, video, or audio) has been manipulated,” with presentation requirements that differ for visual and audio media. The statute also ties coverage to specified election timing and candidate depictions. Teams should read the current codified text and any effective amendments rather than copying that sentence into every national spot.

For each destination, record the statute or agency guidance reviewed, its effective status, the covered office, the election period, the definition of synthetic or deceptive media, the exact notice, placement and duration rules, and any exception being relied on. Save a copy or stable citation. State election agencies and the National Conference of State Legislatures can help locate relevant laws, but a tracker is not a substitute for the enacted text or a court order affecting enforcement.

Proposals belong in a separate column. The Federal Communications Commission opened a proceeding in 2024 concerning disclosures for AI-generated content in political advertising on regulated broadcast services. A notice of proposed rulemaking describes a possible rule; it does not itself impose the proposed requirement. Verify whether the FCC has adopted an effective final rule before marking an AI-specific broadcast label as federally enforced.

Get consent for the synthetic performance and its inputs

Consent should identify the person, permitted uses, channels, election or campaign, editing authority, term, and whether the campaign may train or reuse a voice or likeness model. A general appearance release may not clearly authorize cloned speech, especially when the generated delivery can be revised without another recording session.

The 30-second spot needs evidence that the candidate approved both the use of the voice model and the final words attributed to that voice. If an actor supplied reference audio, the file should also show that the actor had authority to license the recording for model creation and political advertising. Vendor terms matter here: a tool’s permission to process uploaded audio does not prove that the uploader owned the underlying rights.

Do not treat consent as a universal cure. Permission from the depicted person does not remove a statutory disclosure, sponsor disclaimer, or publisher policy. The reverse is also true. Adding an AI label does not grant rights to a recording, performance, photograph, or likeness.

Escalate any file that depicts an opponent, election worker, public official, or private person saying or doing something that did not occur. The review may need to account for state election statutes, rights of publicity, defamation, copyright, fraud, or other laws whose application depends on context. That is a different inquiry from whether a model provider allowed the prompt.

Treat each publisher as another enforcement layer

Ask the station, advertising platform, and social service for their current written election-ad and manipulated-media policies. Private policies can cover content that election law permits, and enforcement may include rejection, reduced distribution, removal, account restrictions, or a platform-applied label.

Provide the final file, payer identity, authorization status, targeting geography, run dates, required certifications, and contact responsible for corrections. Do not assume a platform’s automated label will satisfy a statute. Detection can miss generated material, and a platform label may use different wording or disappear when the file is embedded elsewhere.

Broadcast distribution adds existing FCC obligations unrelated to a proposed AI rule. Stations and other regulated outlets maintain political-file records for covered requests and purchases, including information about the sponsor, schedule, and rates. The publisher may keep the official file, but the campaign should retain what it submitted and reconcile any differences before airtime.

Telephone delivery needs its own gate. The FCC has said AI-generated voices count as “artificial or prerecorded voice” under the Telephone Consumer Protection Act. Political calls using such voices to wireless numbers generally require prior express consent, subject to the applicable rules and exemptions, and prerecorded messages carry identification requirements. Approval for a television spot does not authorize a robocall campaign.

Keep the release packet after the ad stops running

The final packet should connect the approved 30-second export to its script, source rights, synthetic-media assessment, sponsor disclaimer, state analysis, publisher certifications, targeting, invoices, placement records, and any corrected versions. Restrict edit access, but make retrieval possible without depending on the employee who ran the generator.

Retention periods differ. FEC recordkeeping rules generally require political committees to preserve records supporting reports for three years, while FCC political-file rules commonly require covered records to remain available for two years. Other federal, state, contractual, litigation-hold, or campaign policies may demand longer retention, so the packet should name the controlling period rather than default to the shortest one.

If the campaign changes one line after approval, generate a new hash and reopen the affected gates. A revised voice track may alter a required spoken disclosure or exceed a station’s slot even when the video is unchanged. The release decision belongs to the exported file, not to the project name.

Questions people ask

Does every

AI-generated election ad need an AI label?

No blanket federal label covers every format and placement. A state law, effective federal rule, publisher policy, or the facts of a deceptive depiction may require one, while ordinary FEC sponsor disclaimers can apply independently. Check the current rule for every state and channel receiving the file.

Is a

“Paid for by” disclaimer enough for a synthetic ad?

Not necessarily. The FEC disclaimer identifies the payer and, when required, the communication’s authorization status; a synthetic-media notice addresses generated or manipulated content. If both apply, the final export must satisfy the wording, placement, readability, and audibility rules for each.

Can a campaign clone its own candidate’s voice with consent?

Consent resolves only part of the review. The campaign should document permission to create and use the model, approval of the final script, rights in source recordings, and allowed channels, then separately test election disclosures, publisher rules, and telephone restrictions.

What evidence should a publisher request before accepting the ad?

Request the final file identifier, payer and authorization information, targeted jurisdictions, required disclosures, synthetic-media description, rights or consent confirmation, and a responsible contact. Preserve the submitted version and related order records so the published file can be matched to the campaign’s approval packet.