The Disclosure Illusion: Why Labeling AI Campaign Ads Will Not Save Us

Executive summary

• The numbers: the Wesleyan Media Project has identified 220 AI-made political ads in 34 states this cycle, with more than $137 million behind them. In 2024 it found less than $50,000. Only 27% include a disclaimer.
• The policy reflex: around 30 states now require AI disclosure, and the public debate treats the label as the solution.
• The problem: experimental evidence shows labels change what people believe about who made a message, not how persuasive it is. Labels also penalize the candidates who disclose.
• The implication: the real question is not whether an ad is synthetic, but whether it is designed to be judged as a record of reality.

A market that grew in one cycle

The figures reported this month are not a trend line. They are a step change. According to the Wesleyan Media Project, AI-generated advertising went from a rounding error in 2024 to a nine-figure category in 2026 (Gil, 2026). Nearly half of the ads identified feature deepfake depictions of real politicians. One gubernatorial campaign in New York has posted at least 22 of them since April.

The legal system is starting to react, but unevenly. In Texas, a candidate for lieutenant governor filed a criminal complaint under the state's deepfake election law. In Wisconsin and North Carolina, candidates have answered with cease-and-desist letters, in one case in a state with no law on the matter at all (Walker, 2026). Congress has set no federal standard.

In that vacuum, disclosure has become the default answer. Put a label on it, and the voter can decide. It is an attractive idea, because it seems to respect both free expression and the voter's judgment. The evidence says it does neither as well as we assume.

What the label actually does

It does not reduce persuasion

In a randomized experiment with roughly 1,500 U.S. adults, Gallegos and colleagues (2026) showed participants policy messages generated by a large language model. Some were told an AI wrote them, some were told a human expert did, and some were told nothing. The labels worked in the narrow sense: people updated their beliefs about authorship. They did not work in the sense that matters. The messages were equally persuasive across conditions, across topics and across demographic groups.

The finding is uncomfortable but intuitive. Persuasion runs on content, emotion and repetition, not on metadata. A voter who sees a convincing scene and a small line of text saying "AI-generated" still carries the scene home. The Axios reporting makes the same point from the field: experts note that deepfakes can be effective even when viewers know they are fake (Walker, 2026).

It punishes the honest

The second problem is more perverse. A study by the American Association of Political Consultants Foundation found a "disclaimer effect": when viewers saw an AI disclosure, trust in the message dropped sharply, even when the ad contained no AI content at all (Campaigns & Elections, 2026). Larger labels produced larger penalties.

Combine the two findings and the incentive is clear. A campaign that complies pays a credibility cost. A campaign that ignores the rule, or claims its ad is "parody," pays nothing until a regulator or a court catches up, usually after the election. With 73% of ads already undisclosed, the market has done the math.

It makes the unlabeled look true

There is a third effect, documented before generative AI existed. Pennycook, Bear, Collins and Rand (2020) showed that when warnings are attached to some false headlines, people rate the unwarned false headlines as more accurate. They called it the implied truth effect. A labeling regime that catches a quarter of synthetic ads may quietly certify the other three quarters.

The older diagnosis

None of this would have surprised Daniel Boorstin. In The Image (1961/1992), he described the "pseudo-event": an occurrence planned primarily to be reported, whose relationship to reality is deliberately ambiguous. The press conference, the staged visit and the photo opportunity are pseudo-events. Their power, Boorstin argued, comes precisely from the fact that audiences know they are staged and consume them anyway.

The AI campaign ad is the pseudo-event in its purest form. It no longer needs an event at all. And Boorstin's lesson applies directly: knowing that something is manufactured has never been enough to neutralize it. A label tells the audience what they already half-know. It does not restore the distinction between a representation and a record.

That is why the more useful line is not between synthetic and authentic, but between two kinds of claim an ad can make. An ad can be expressive, an obvious caricature, a cartoon or a metaphor. Or it can be testimonial, presented as evidence that a person said or did something. The harm lies in the second, whether or not AI was involved.

What leaders and campaigns should take from this

1. Do not outsource integrity to the label. Compliance with disclosure laws is necessary, but it is not a reputational strategy. Assume a disclosed ad will be judged harshly and an undisclosed one will be discovered.
2. Draw your own line at testimony. Never put words or actions in a real person's mouth, labeled or not. That is the standard that will survive litigation and public scrutiny.
3. Treat authenticity as an asset. One Texas Senate candidate now ends his ads with "Made by humans. Not AI." (Gil, 2026). Verifiable presence, real footage, named witnesses and live events are becoming a competitive advantage, not a nostalgic one.
4. Prepare the counter-record in advance. If your organization may be the target of synthetic testimony, the response cannot start with a denial. It must start with documentation that already exists: timestamps, recordings and third parties who can vouch for what happened.

The disclosure debate assumes the problem is information: if voters only knew, they would discount. The evidence suggests the problem is judgment, and judgment is shaped long before the label appears.

References

Boorstin, D. J. (1992). The image: A guide to pseudo-events in America. Vintage Books. (Original work published 1961)

Campaigns & Elections. (2026, April 10). AI disclaimers on campaign ads are sowing mistrust, study finds. https://campaignsandelections.com/industry-news/ai-disclaimers-on-campaign-ads-are-sowing-mistrust/

Gallegos, I., Shani, C., Shi, W., Bianchi, F., Gainsburg, I. B., Jurafsky, D., & Willer, R. (2026). Labeling messages as AI-generated does not reduce their persuasive effects. PNAS Nexus, pgag008. https://doi.org/10.1093/pnasnexus/pgag008

Gil, B. (2026, October 16). $137 million has been spent on AI-generated ads ahead of midterms. Only 27% included a disclaimer. Gizmodo. https://gizmodo.com/137-million-has-been-spent-on-ai-generated-ads-ahead-of-midterms-only-27-included-a-disclaimer-2000827138

Pennycook, G., Bear, A., Collins, E. T., & Rand, D. G. (2020). The implied truth effect: Attaching warnings to a subset of fake news headlines increases perceived accuracy of headlines without warnings. Management Science, 66(11), 4944–4957. https://doi.org/10.1287/mnsc.2019.3478

Walker, J. (2026, October 2). The AI deepfake election is starting to get litigious. Axios. https://www.axios.com/2026/10/02/ai-deepfake-election-campaign-lawsuits