The AI Ad Disclaimer That Put the Sender in Charge of Warning the Voter

A safety rule that depends on the party with the most to lose from complying is not a safeguard. It is a request.

Tuesday’s midterms were, in the words of the Wesleyan Media Project’s researchers, the first in which AI-generated political ads regularly appeared on televisions and social media feeds. The project has tracked political advertising since 2010, and its running count tells the story. In May it had found 33 AI-made or AI-enhanced ads and about $20 million in spending. By early September it had 164 ads and nearly $80 million. Through October 11, the latest tally I found, it counted 220 ads in 34 states and more than $137 million. In 2024, it could identify less than $50,000. The project says its count is likely an undercount, because AI use is often impossible to prove, and the tally stops before the final weeks, when ad buying is heaviest.

The ads are not confined to one side. AFP reported that both parties’ campaigns were using deepfake ads to target rivals, including a video depicting a Senate candidate laughing at cards reading “Death to America” and “Celebrate 9/11.” Most of the ads the project identified, more than 80%, came from Republican candidates and allied groups, with Democrats about 12%. That describes the ads researchers could detect and identify, which isn’t necessarily all the ads that ran.

The legal backdrop is a patchwork. No federal law specifically regulates AI in political ads. The National Conference of State Legislatures counted 31 states with election deepfake laws as of June. Four ban certain uses of deepfakes, and 27 others require some form of disclosure.

Then came the disclosure numbers. Wesleyan’s co-director Erika Franklin Fowler said only about 27% of the identified ads carried an AI disclaimer, and called the trend problematic. In an AP report on the project’s earlier tally of 164 ads, co-director Travis Ridout cross-referenced the data with state laws. In states with no AI disclosure law, 32% of the AI ads carried a disclaimer anyway. In states with regulations, 29% did. A separate CNN analysis found at least $38 million went to ads that appear to use AI imagery without disclosure.

Roughly seven in ten identified AI ads ran without an AI disclaimer, and states that required one did no better than states that didn’t.

The laws were largely written for a different threat. Many followed the January 2024 New Hampshire robocall, in which a cloned voice resembling then-President Biden urged Democrats not to vote. That was anonymous impersonation designed to suppress turnout. What arrived in 2026 looks different: funded, full-length, candidate-linked ads, some openly styled as parody or enhancement, in a legal landscape built around disclosure rather than prohibition. Fowler said even her team was surprised by how AI was used, and that without a disclaimer, some uses would be hard to detect.

The pattern is familiar from other transparency rules. A requirement asks the party with the most to lose to supply the warning, and compliance becomes a design decision.

The limits of transparency: Requiring disclosure doesn’t guarantee audiences are protected. Disclosers may feel licensed to be more self-serving once they have warned, and audiences often don’t adjust enough for what was disclosed (Loewenstein, Cain and Sah, 2011).

An AI label costs the sponsor and benefits the viewer. Evidence on the size of that cost is mixed, but its direction is consistent.

Disclaimer effect: The finding, in a study run by the American Association of Political Consultants Foundation, that an AI disclosure reduced viewers’ approval of an ad’s message even when the ad used no AI.

That study is industry-run, and consultants have a stake in fewer disclaimers, so treat its size with caution. A 2024 experiment reported by Marketplace found AI labels largely hurt candidates across the board, for harmless and deceptive uses alike, and hurt the candidate running an attack ad more than the one attacked. Effects were small, partly because many viewers didn’t notice the labels. A small cost still shapes behavior. If a label lowers approval, the sponsor’s best move is the least visible compliant label, or none where the law allows.

The label may also fail at its job.

Implied truth effect: When only some content carries a warning, content without one is judged more credible than it deserves, because people read the absence of a warning as a clean bill of health (Pennycook, Bear, Collins and Rand, 2020).

With roughly a quarter of identified AI ads labeled, a video without a label may look more authentic than it is. That is my inference by analogy, not a measured result in political ads.

Finally, labels may not touch persuasion at all. In a pre-registered experiment of 1,601 Americans, Gallegos and colleagues at Stanford varied whether a policy message was labeled as written by AI, written by a human expert, or unlabeled. The label had no significant effect on persuasion, across party and prior AI experience. That study tested text, not video, but it points the same way. Labels address who made a message, while persuasion tends to follow what the message says.

For campaigns, the incentive is plain: a rule that makes the sponsor pay for visibility invites the smallest compliant label. The AP comparison fits that pattern, though it’s small and confounded.

For voters, partial labeling can teach the wrong lesson. If some ads are labeled and most aren’t, “no label” starts to mean “real.”

For regulators, enforcement runs through complaints, which means it arrives after the ad is on the air. Laws written after one impersonation robocall are now being tested by a market for funded synthetic ads.

The lesson applies to any disclosure regime, not just political ads.

• Move the warning away from the party being warned about. Provenance at the source, or platform-level labeling, doesn’t depend on the sponsor’s willingness. Public Citizen has argued for labels as clear as food labels.

• Standardize format, size, duration, and wording. The 2024 experiment found design and wording matter, and many viewers never noticed the labels.

• Label what was fabricated, not just that AI was used. The studies above found labels penalized harmless enhancement and deceptive fabrication alike. A label saying a depiction of a real person is synthetic and shows events that did not occur is more informative than a generic AI tag.

• Pair labels with fast remedies for fabricated depictions of real candidates. Labels alone don’t prevent harm in the days an ad is running. Four states already ban certain uses.

• Test whether viewers notice and understand a label before mandating it. The research so far measures persuasion and approval. It rarely measures whether viewers believe depicted events happened.

• Publish compliance rates by state. Independent trackers like Wesleyan make noncompliance visible, which changes the sponsor’s cost-benefit math.

• The data can’t show labels fail to protect voters. No study here tests whether a disclosure law changes how people vote, and the tallies are undercounts that end on October 11. Disclaimer rates also vary across tallies, from 27% to 31%.

• The state comparison is small and confounded. The 29% versus 32% gap comes from an early tally of 164 ads, and states differ in laws, races, and ad types.

• The studies have limits. The Stanford experiment tested text, not video. The AAPC study is industry-run. The 2024 experiment used fictional ads.

• Political ads generally move votes less than people assume. The stakes here may be about trust and misinformation more than vote switching, and I can’t attribute any result to AI ads.

• The partisan split describes identified ads. It may reflect detection and who produced detectable ads, and it supports no conclusion about either party’s intent.

• Disclosure still has value. Even imperfect labels give journalists and watchdogs something to check and help establish norms. Free speech protections also constrain what states can ban, including parody.

In your own industry, who pays the cost of a disclosure, and who gets the benefit?