A political ad can put authentic words in a candidate’s mouth while inventing the performance around them. For campaigns and broadcasters, the compliance problem is that an AI label explains how an ad was made without necessarily helping viewers work out which parts to doubt.
A Fabricated Performance Creates a Real Review Problem
An 85-second attack ad targeting Texas Senate candidate James Talarico illustrates the challenge. Produced by the Senate Republicans’ campaign arm, it combines readings of his past tweets with synthetic commentary he did not deliver. A faint disclosure appears in a corner of the ad. UC Berkeley deepfake expert Hany Farid judged that viewers would struggle to spot the fabrication right away.
Checking the underlying tweets would address only part of an advertising reviewer’s task. The presentation also implies that the candidate appeared on camera and delivered the entire message. Accurate source material can therefore sit inside a misleading depiction.
The Wesleyan Media Project reported that spending to air AI-made political advertising rose from under $50,000 across the 2024 cycle to more than $137 million during the current cycle. The totals cover AI advertising broadly, rather than deceptive content alone. Even with that distinction, the rise calls for advertising operations to make synthetic-media review part of their routine work.
Production Methods Are a Poor Substitute for Risk Categories
Campaigns use generative tools for very different purposes. Animating a photograph or generating a household scene does not create the same concern as manufacturing evidence of a candidate’s conduct.
The practical attraction is clear. Fred Davis, who has made campaign ads for 50 years, estimated that a scene costing $2 million to produce could cost about $200,000 if generated instead. Those savings help explain why campaigns adopt the tools, but judging whether a depiction is acceptable takes a separate review.
A useful review process would separate production choices from claims about real people:
| Type of depiction | Main review question |
|---|---|
| Cosmetic edits or synthetic background scenes | Does the alteration change what viewers are being asked to believe? |
| Obviously fantastical political satire | Could the presentation reasonably be mistaken for an actual event? |
| Realistic depictions of identifiable people | Does it invent speech, behavior, or an association that audiences may accept as genuine? |
These are proposed review categories, not a description of existing legal exemptions. Their value is in directing attention toward the effect of an ad instead of treating every use of AI as equally risky.
That distinction becomes concrete in an ad from Make Oklahoma Great Again PAC. It depicted gubernatorial candidate Mike Mazzei alongside Hillary Clinton, carried no AI disclosure, and received more than $200,000 in Oklahoma television spending. The governance concern is the apparent association created by the image, regardless of how sophisticated the fabrication looks.
Existing Disclosure Rules Leave an Implementation Gap
At least 31 states have election-related deepfake laws. Yet Wesleyan’s researchers found disclosures on 29% of AI advertisements in states with such laws and 32% in states without them.
That comparison does not establish why individual campaigns omitted labels or whether each ad breached a requirement. It does show that the presence of legislation alone is a weak measure of effective disclosure.
Wording and visibility create a second problem. In research involving 1,051 participants, NYU’s Center on Technology Policy found that 37% did not recall Michigan’s required disclaimer. Only 21% understood that the wording referred to AI. A disclosure can satisfy a prescribed format while communicating little to its intended audience.
Campaign incentives also complicate implementation. The research found that labels reduced trust in the candidate sponsoring the advertisement. Where disclosure carries a reputational cost, oversight needs to address omissions rather than assume sponsors will volunteer clear warnings.
Consent Could Provide a More Workable Screening Test
One proposed approach would make permission from the depicted person the trigger for disclosure. Without that permission, a paid political advertisement would need an understandable warning in both its visuals and audio before being aired.
This would shift review away from an argument over whether a particular production tool counts as AI. Broadcasters would then check what the advertisement depicts and whether the person approved that portrayal, along with any warning the ad is required to carry.
A campaign might, for example, submit an attack ad that uses a synthetic version of its opponent to deliver invented remarks. Checking the script’s factual references alone wouldn’t complete the review under this approach. The submission process would also need documentation of consent or the appropriate disclosure, with responsibility assigned for checking both before release.
The proposal has a deliberately limited scope: paid political advertising. Extending the same workflow to private messages or satire accounts would raise different practical questions.
Federal Consistency Would Still Require Operational Controls
H.R. 9922 would establish a federal disclaimer requirement. Political consultants’ trade representatives have also sought a national standard in place of differing state approaches. A common framework could simplify review, although consistent rules would still need clear ownership inside campaigns and advertising operations.
Sejin Paik, the Georgetown researcher who now leads TrueMedia, has emphasized television’s particular limitation: viewers have little opportunity for surrounding reactions or provenance information to correct a deceptive impression as an ad plays.
That makes pre-air review especially consequential. Preparing for the 2028 presidential election means deciding who approves synthetic portrayals, what evidence must accompany submissions, and how missing disclosures will be resolved. A warning requirement can work only if production and placement processes reliably get the warning into the ad.
