How US political ad deepfakes meet FEC disclaimers, platform policies, and C2PA checks

What to take away

What counts as a deepfake in a US political ad

A deepfake in a political ad is synthetic or manipulated audio, video, or image content that depicts a real person saying or doing something they did not say or do. The term covers face swaps, voice cloning, lip-sync edits, and generated video built from a candidate's public appearances.

Background on the technology and its detection is collected in Deepfake - Wikipedia.

The US has no single federal deepfake statute aimed at campaigns. Instead, a political ad can run into several overlapping regimes: FEC disclosure and fraud rules, Federal Trade Commission authority over deceptive advertising, state election statutes, and platform terms of service.

That patchwork matters because a video can be lawful in one state and actionable in another.

Context decides whether a clip is treated as satire, commentary, or fraud. A clearly labeled parody that no reasonable viewer would take literally is treated differently from a fabricated endorsement released days before an election. Ad reviewers should ask three questions: who is depicted, what action is implied, and what a typical viewer would believe.

Manipulated media is not always fully synthetic. Cheap fakes, such as slowed audio or trimmed video, can mislead as effectively as generated footage. The detection habits that work for generated content also work here: check the original source, the upload history, and whether any independent camera captured the same moment.

FEC ad disclaimers and what they do not cover

FEC ad disclaimers are the notices that state who paid for a communication and whether it was authorized by a candidate. They appear as spoken lines, on-screen text, or both, depending on the medium and the audience size. The commission publishes Guidance documents | FEC that committees use to get the wording and placement right.

Disclaimers answer a funding question, not a truth question. A paid-for notice tells you a committee stands behind the ad. It does not tell you the footage is authentic, that the audio was not cloned, or that a depicted event occurred. Voters who treat a disclaimer as a verification stamp are reading it for more than it says.

The disclaimer rules also vary by channel. Television and radio spots carry spoken disclaimers, while small digital ads can use a landing page or hover text in some circumstances. A synthetic video can satisfy every disclaimer requirement and still be a fabrication, which is why ad reviewers need a separate authenticity check.

For committees, the practical risk is not only a fine. A deceptive ad can trigger platform removal, state enforcement, and defamation exposure if it falsely attributes statements to a real person. The safe pattern is to label synthetic elements clearly, keep production records, and preserve the source files behind any manipulated footage.

FEC advisory opinions and guidance on synthetic media disclosure

FEC advisory opinions are formal responses to specific questions from campaigns, committees, and outside groups. They are the agency's clearest public signal on how existing rules apply to new tactics, including synthetic content. The full set is searchable at Advisory opinions | FEC.

The clearest recent marker on synthetic media is a petition, not a decided opinion. Public Citizen asked the commission in July 2023 to write rules on AI-generated campaign content. The commission declined to open that rulemaking in 2024.

That decision leaves advisory opinions as the route for a committee with a specific question. A request gets a 60-day response window under the commission's regulations, with a possible extension. Each published opinion lists the requester, the date, and the outcome, and the FEC posts the request beside the response.

An advisory opinion binds only the requester and the facts described, but it shapes how other committees behave. If a group asks whether a synthetic voice in a fundraising appeal needs a specific disclaimer, the answer becomes a reference point for similar ads. Reading the request alongside the response shows which facts the commission treated as decisive.

Requests take time, and election calendars do not wait. That gap is why many committees rely on published guidance, prior opinions, and their own counsel rather than filing a fresh request. When a question is genuinely novel, the request itself can become news, and reporters should distinguish a pending request from a decided opinion.

Advisory opinions do not create new disclosure duties for synthetic media by themselves. They interpret the statute and rules already on the books. A committee that wants certainty about labeling synthetic video should expect to describe its production process, its disclaimers, and how viewers will encounter the ad.

Platform synthetic media policies compared

Platform synthetic media policies are the private rules that govern synthetic and manipulated political content. They are not law, and they change without notice. A campaign that plans a synthetic ad should read the current policy for each service it intends to use and keep a dated copy.

These are summaries of published rules, not the full policy text, and the wording changes. Some services require disclosure in the upload flow, others rely on detection, and several use both. Enforcement is inconsistent, so a removed ad is not proof of falsity and a surviving ad is not proof of authenticity.

Appeals matter for campaign staff. When a synthetic ad is pulled, the committee usually has a short window to contest the decision and may need to supply source files or a production statement. Keeping a clean chain of custody for every generated asset makes that appeal faster and gives reviewers something concrete to evaluate.

State law adds another layer. California AB 2839, signed in 2024, amended Elections Code section 20010, which lets candidates and others sue over deceptive AI-generated election content. That section began as AB 730 in 2019, which criminalized deepfake audio or video of a candidate within 60 days of an election.

Texas Election Code section 255.004 bars distributing a deepfake video of a candidate within 30 days of an election. Washington SB 5152, passed in 2023, requires disclosure when synthetic media appears in electioneering communications. Parts of the California law face a federal court challenge, so campaigns should confirm its current status.

C2PA provenance signals and content credentials

C2PA provenance signals are metadata that record how a file was captured and edited. The standard, developed by the Coalition for Content Provenance and Authenticity, attaches a signed content credential that can list the camera, the editing steps, and any synthetic tools used. A viewer with a compatible tool can inspect that history.

Credentials travel with the file only if platforms and editors preserve them. Social services often strip metadata on upload, so a missing credential is common and does not mean a video is fake. The useful signal is a present, valid credential that contradicts the claim being made, or a credential that shows a synthetic step the poster denied.

For ad reviewers, provenance is strongest when it matches an independent record. If a credential says a clip came from a specific camera at a specific time, that claim can be checked against the original broadcast or a second camera angle. Provenance alone is a claim about a file, not a guarantee about the event.

Provenance also helps with the opposite problem: proving that authentic footage is real. A campaign that publishes raw files with credentials intact gives fact-checkers a way to confirm a clip quickly. That is a defensive habit worth building before a controversy, not after one.

Testing detection tools against NIST AI benchmarks

Detection tools for synthetic media range from free web uploads to enterprise systems. Their accuracy varies by generator, medium, and compression, so a tool that catches one face-swap technique may miss a cloned voice. NIST publishes standards and research relevant to these systems at Artificial intelligence | NIST.

NIST AI 600-1, the Generative AI Profile issued in July 2024, sets out risk categories for synthetic content, including information integrity and confabulation. NIST's GenAI program has run evaluation challenges, including one that asks participants to tell machine-written text from human text.

Those tests cover text, and no NIST benchmark yet scores video deepfake detectors on political advertising. For faces, NIST runs the Face Recognition Vendor Test and the Face Analysis Technology Evaluation, which report error rates across demographic groups. That work shows why a single accuracy figure hides more than it settles.

Benchmarks give reviewers a way to compare tools on the same footing. A benchmark defines test data, tasks, and scoring, which exposes whether a detector was evaluated on realistic political content or only on clean laboratory samples. Without that context, an accuracy figure is close to meaningless.

When you test a detector, use material that resembles your actual workload: compressed social video, short clips, low light, and speech from real public figures. Record the false positive and false negative rates you observe. A tool that flags every authentic clip is unusable even if it catches every fake.

Audio deserves its own pass. Most political deepfake coverage focuses on video, but cloned voices spread through phone calls, robocalls, and short audio clips where viewers have no visual cues to weigh. If a story rests on a recording, get a clean copy of the original file and compare it against known samples of the same speaker.

The evidence types that carry the most weight are original files, independent captures, and consistent witness accounts, and our image verification checklist sets out how to work through them.

Reading a detector card before trusting its verdict

The card that ships with a detection tool states what it was built on, how it was tested, and where it fails. Read it for development data, known limitations, and the intended use. Many tools are built for one medium and perform poorly outside it.

Look at which generator families appear in the development data. A detector built before a new video generator existed may not recognize its output. Check the languages and accents covered, since voice detectors often degrade on speech outside the material they were built around.

The AI Risk Management Framework | NIST offers a structured way to assess these risks. It asks teams to map context, measure performance, and manage uncertainty, which is exactly what an ad reviewer needs when a vendor ships no evaluation record.

These cards also reveal governance. A vendor that documents evaluation, versioning, and error reporting is easier to defend in a newsroom or a compliance review. A vendor that publishes only a marketing page leaves you carrying the risk alone.

Common synthetic media reporting problems in campaign coverage

Reporting on synthetic campaign video fails in predictable ways. The most common is treating a detector score as proof. A percentage from a web tool is not evidence of fabrication, and publishing it as such invites a correction. Our guide to image and video verification walks through the recurring problems that outlets run into before deadline.

Another failure is ignoring the original upload. Reporters sometimes analyze a recompressed clip from a social repost instead of the earliest available file. That destroys provenance and makes the analysis harder to defend.

A third problem is missing the disclosure question. Coverage often focuses on whether a video is fake and skips whether the ad carried the required FEC ad disclaimers or a platform label. Both matter to viewers trying to judge what they are seeing.

Use a short checklist before publishing any synthetic media story:

When a suspected fake involves audio, the stakes rise because voice is harder for viewers to judge. Our account verification checklist shows how quickly a fabricated recording can spread before verification catches up.

A repeatable process beats a one-off scramble. Start by learning how to verify a social media account, then move to a full method for handling social verification problems. The steps below reflect that order.

Common questions