Wesleyan AI Campaign Ads Data Exposes a $137 Million Disclosure Gap

Wesleyan AI campaign ads data has identified more than 220 synthetic or AI-enhanced videos promoted with over $137 million during the 2026 midterms.

About 70 percent carried no disclosure that artificial intelligence was involved. More than 80 percent came from Republican campaigns or Republican-aligned political action committees.

Those figures mark a change in how generative media operates in American politics. AI is no longer limited to experimental posts or easily dismissed internet jokes.

Campaigns are now paying to place synthetic images on television and digital platforms. Some ads make subtle visual adjustments, while others fabricate scenes involving recognizable politicians.

The central conflict is no longer simply whether political deepfakes exist. It is whether voters can recognize synthetic persuasion before it shapes their view of a candidate.

Federal rules generally tell voters who purchased an advertisement. They do not provide a universal requirement identifying how its images, voices, or scenes were produced.

That distinction has created a disclosure gap at the center of the 2026 election. Campaign attribution can be present even when the ad’s synthetic construction remains hidden.

The Wesleyan AI Campaign Ads Count Has Surged

The important change is not one viral deepfake, but the arrival of synthetic media as a funded campaign advertising category.

The midterm ad analysis published by The New York Times cited research from the Wesleyan Media Project. The researchers identified more than 220 videos using AI-generated or AI-enhanced material.

Candidates and political groups spent more than $137 million promoting those ads during 2026. The spending estimate covers distribution, not the cost of generating the underlying images or video.

That distinction matters. Generative tools can make an individual scene relatively easy to produce, but campaign spending determines how often voters encounter it.

The dataset also points to a large partisan imbalance. More than 80 percent of the identified ads came from Republican candidates or Republican-aligned political groups.

Republicans have therefore become the leading adopters during this cycle. Democratic campaigns have also used AI, but at a much smaller observed scale.

The current totals represent a sharp expansion from earlier findings. In September, the project’s AI advertising dataset contained 164 ads associated with nearly $80 million in estimated spending.

That September release found that 80 percent of identified ads came from Republican or pro-Republican sponsors. Those sponsors accounted for 83 percent of tracked spending.

Eleven percent of the ads came from pro-Democratic sponsors. Other material supported independent candidates, issue campaigns, or sponsors outside the two principal partisan categories.

The disclosure findings were already striking. Sixty-nine percent of those 164 ads lacked an AI notice, while undisclosed ads represented 78 percent of spending.

The later count raised the observed total beyond 220 and the associated spending above $137 million. Yet the proportion without labels remained near 70 percent.

The Wesleyan Media Project builds its list from several signals. Researchers follow press reports, search for AI disclaimers, monitor advertising, and review suspected synthetic visuals.

They also use information from AdImpact, an advertising tracking company, to estimate spending. Confirmed cases range from enhanced still images to fabricated depictions of real people.

Researchers describe their numbers as an undercount. AI-assisted editing can be difficult to detect, especially when an ad contains no label and avoids obvious visual errors.

That limitation makes the $137 million figure a floor rather than a complete market total. It measures known AI campaign ads, not every political advertisement touched by AI.

The category itself is also broad. An ad with an invented background and an ad that fabricates a candidate’s speech both qualify for the dataset.

That breadth does not make the count meaningless. Instead, it reveals why a single disclosure rule is difficult to design.

Voters face content ranging from ordinary production assistance to false performances by public figures. Both involve AI, but they create different levels of risk.

The 2026 cycle has turned that classification problem into an immediate policy dispute. Millions of dollars are already supporting the content before regulators agree on common definitions.

Synthetic Media Now Covers Editing, Satire, and Impersonation

AI campaign ads are difficult to regulate because the technology supports both familiar editing practices and convincing factual fabrications.

One end of the spectrum looks like conventional campaign production. AI can remove objects, generate backgrounds, extend footage, or visualize events for which no photograph exists.

Supporters of these uses argue that the technology functions like another editing tool. They say a broad label can imply deception even when an ad preserves the underlying claim.

Political consultant Brad Todd discussed one such scenario in The New York Times report. AI might recreate an emotional childhood scene when no historical image exists.

That reconstruction can express a real biographical memory without pretending that archival footage survived. However, a viewer may still interpret the generated scene as documentary evidence.

At the other end are deepfakes, meaning synthetic media that depicts a recognizable person saying or doing something that never occurred.

Some 2026 ads have placed candidates beside polarizing national figures. Others have generated speech, altered physical appearances, or invented provocative behavior.

An Oklahoma advertisement depicted a Democratic gubernatorial candidate embracing Hillary Clinton. The group behind it reportedly spent more than $200,000 promoting the unlabeled video on television.

Other ads have used visibly absurd imagery. A Michigan Senate campaign depicted Republican Mike Rogers performing impossible superhero feats, including stopping a moving car.

A North Carolina advertisement showed Democratic candidate Jessica Bichler twisting into physically impossible yoga positions. That video openly acknowledged its artificial construction.

Those examples resemble political cartoons translated into moving images. Their exaggeration gives viewers contextual clues, even when their precise production method is unfamiliar.

More realistic advertisements create a harder problem. Bruce Blakeman’s campaign in New York generated depictions of Gov. Kathy Hochul and Mayor Zohran Mamdani walking together.

The video identified Blakeman’s campaign as the sponsor. However, it did not identify the depictions of Hochul and Mamdani as artificially generated.

A sponsor label answers who paid for the message. An AI label answers whether the events shown actually occurred.

That difference is essential. A viewer can understand an ad’s partisan origin while still believing that its visual evidence came from reality.

Blakeman’s campaign produced at least 22 AI-generated advertisements beginning in April, according to the Times analysis. Only a small portion included AI or manipulation disclosures.

The New York State Democratic Party filed a complaint concerning his campaign’s use of AI. It argued that materially deceptive political communications required a label.

Blakeman’s campaign reportedly described one disputed advertisement as satire. That defense illustrates the boundary regulators must evaluate repeatedly.

Satire depends on audience recognition, context, and exaggeration. Synthetic media can remove those signals while preserving the creator’s ability to claim humorous intent.

The same tension appears in other races. AI-generated attack ads have targeted Democratic candidates James Talarico, Abdul El-Sayed, Josh Shapiro, and Zach Dembo.

Some fabrications are plainly theatrical. Others place words or behavior directly onto a candidate, making the content resemble an authentic recorded event.

This range explains why “AI-generated” alone is an incomplete warning. The phrase does not tell voters which elements were altered or whether the central claim is factual.

A useful disclosure must distinguish production assistance from material fabrication. Material fabrication changes evidence that a reasonable viewer might use to judge a candidate.

That approach still requires difficult decisions. A generated background may be incidental in one advertisement and central to the emotional argument in another.

Campaigns also have incentives to choose the narrowest interpretation. A prominent AI label can reduce trust even when the depicted political claim remains accurate.

Generative media therefore creates a strategic advantage beyond lower production costs. It lets campaigns construct scenes that never existed while debating whether those scenes require disclosure.

Republican Adoption Is Pressuring Campaign Norms

Republican campaigns have established the cycle’s dominant AI advertising strategy, forcing opponents and consultants to reconsider their own limits.

The partisan imbalance is one of the strongest findings in the Wesleyan AI campaign ads research. It is not merely a difference in social posting styles.

Republican candidates and aligned groups accounted for most identified advertisements and most associated spending. Their uses range from small enhancements to full synthetic impersonations.

AI gives campaigns a fast method for connecting local opponents with unpopular national figures. It can create the visual association without arranging footage or finding an existing photograph.

A generated embrace, dinner, protest, or private conversation can compress an argument into several seconds. The image supplies an emotional conclusion before supporting facts appear.

That makes the technology especially suitable for negative advertising. Campaigns can create a memorable accusation even when no real footage illustrates it.

The technique also works across television and online distribution. A synthetic clip can begin on a campaign account, enter paid digital rotation, and later reach broadcast audiences.

Each step gives the image more perceived legitimacy. Repetition can make a fabricated scene familiar even when viewers never consciously accept it as authentic.

Opponents face an uneven response problem. Ignoring the content lets it circulate, while condemning it can expand its reach.

Producing a rebuttal also requires time. Generative tools can create new provocative scenes faster than campaigns, journalists, or regulators can investigate each one.

Democratic campaigns must now decide whether to adopt similar production methods. Refusing them can preserve a distinction, but it can also surrender attention.

Using the same methods carries separate risks. It can weaken criticism of Republican deepfakes and accelerate an advertising norm with fewer boundaries.

The conflict is therefore not simply Republicans against Democrats. The central contest is aggressive synthetic persuasion against meaningful disclosure.

Republican campaigns currently represent the clearest adopters of the aggressive approach. Yet the underlying incentives are available to every party and outside group.

Political action committees deserve particular attention because they can operate independently from candidates. Their advertisements may use harsher content than a candidate wants to authorize directly.

Federal campaign rules already require many independent communications to identify their financial sponsor. They also require a statement that the candidate did not authorize the message.

Those requirements create accountability for payment. They do not automatically explain whether a person’s appearance, voice, or action was generated.

The federal disclaimer rules set detailed standards for television, radio, print, and internet communications.

Television disclaimers must meet requirements for readability, duration, and size. Candidate-authorized broadcast advertisements also include an approval statement.

Internet public communications generally need visible audio or visual attribution. Yet those rules concern sponsorship rather than comprehensive AI provenance.

Provenance describes where media came from and how it changed. For political advertising, it can include the source footage, editing history, model use, and responsible publisher.

An ad can comply with sponsor attribution while leaving its media provenance unclear. The Blakeman example demonstrates how those two disclosures diverge.

Campaign consultants also have a commercial reason to resist broad labeling. A label can make harmless editing appear equivalent to a deceptive candidate impersonation.

Julie Sweet of the American Association of Political Consultants has argued for a federal standard focused on deceptive synthetic content.

That position acknowledges the risk without treating every automated production step alike. It also reflects frustration with different state definitions and deadlines.

A uniform standard could reduce compliance confusion. However, a narrow definition might leave consequential emotional manipulation outside the rule.

The political advantage currently lies with campaigns willing to test those boundaries. Their opponents must respond before courts and election agencies establish reliable precedent.

The Disclosure Patchwork Is Not Producing Consistent Labels

State AI disclosure laws have expanded, but the observed labeling rate does not show that legal coverage guarantees practical transparency.

The Associated Press examined an earlier version of the Wesleyan dataset and found a counterintuitive pattern in the available sample.

Thirty-two percent of ads from states without an applicable AI disclosure law contained a disclaimer. Only 29 percent from regulated states carried one.

Those percentages do not prove that disclosure laws reduce compliance. Wesleyan co-director Travis Ridout emphasized important limitations in the comparison.

Some tracked advertisements fell outside the relevant law’s scope. A state might regulate candidate impersonation but not an invented background or exaggerated physical scene.

Timing can also change whether a law applies. An advertisement may have aired before a new statute took effect.

Still, the disclosure investigation shows why enactment alone is an incomplete measure. Definitions, enforcement, and campaign guidance determine whether voters see labels.

The state law tracker maintained by the National Conference of State Legislatures reveals wide differences among jurisdictions.

Some states focus on materially deceptive depictions of candidates. Others cover synthetic media intended to influence an election, regardless of demonstrated deception.

Several statutes apply only within a defined period before voting. Others create civil remedies, criminal penalties, or rights to seek an injunction.

Exemptions for satire, parody, news reporting, and platform activity vary as well. Courts have challenged provisions that burden protected political speech too broadly.

California’s rules illustrate the constitutional pressure. Courts objected to language tied to vague electoral harm and to burdensome requirements for satire.

Hawaii’s statute also faced a permanent injunction. These cases show that regulators cannot simply prohibit every synthetic depiction connected to politics.

Federal policy remains narrower. In 2024, the Federal Election Commission declined to open a dedicated AI advertising rulemaking.

Its AI interpretive rule instead clarified that existing fraudulent misrepresentation provisions remain technology neutral.

Those provisions cover specific conduct. A candidate or agent cannot falsely claim to speak or act for another candidate or political party in a damaging way.

They also restrict false representations made while soliciting contributions. AI can facilitate those violations, but the rule does not create a universal labeling mandate.

That leaves many synthetic attack ads in a gray area. They can be misleading without satisfying the narrow elements of federal fraudulent misrepresentation.

The disclosure debate also contains a less obvious complication. Labels do not guarantee that voters notice, understand, or correctly interpret them.

A 2024 AI label experiment tested disclaimers on fictional political advertisements with 1,051 online participants.

Participants generally rated candidates using labeled AI ads as less trustworthy and less appealing. They also viewed the labeled ads as less accurate.

The effect appeared across deceptive and more harmless uses. Attack-ad labels produced a backfire effect against the candidate sponsoring the advertisement.

As many as 37 percent of participants did not remember seeing the AI label. Wording also influenced what viewers thought had been altered.

These findings explain why campaigns may avoid voluntary disclosure. A label can impose a political cost without ensuring that every viewer understands the media.

That incentive does not excuse noncompliance. It does show why legal requirements need clear definitions, prominent presentation, and credible enforcement.

A generic badge can also flatten meaningful distinctions. It may place minor editing and fabricated candidate speech in the same mental category.

The strongest disclosure system would identify the relevant alteration. It could tell viewers that a voice was generated or that the depicted event never occurred.

Such specificity is harder to standardize. It requires creators to maintain records and platforms to preserve provenance through editing, compression, and reposting.

It also requires election officials to act quickly. A successful remedy delivered after voting cannot reverse every impression created during the campaign.

Three Signals Will Show Whether the Gap Narrows

The next test is whether enforcement, campaign behavior, and voter response begin moving faster than synthetic advertising production.

The first signal is the resolution of complaints under existing state laws. The New York complaint involving Bruce Blakeman offers a visible early test.

A clear ruling would help define when satire becomes materially deceptive synthetic media. It would also show whether regulators can act before an election ends.

If authorities require prominent corrections or labels, the pressure for voluntary disclosure will increase. A slow or inconclusive process will strengthen the current incentive to publish first.

The second signal is the final ratio of labeled to unlabeled advertisements. The Wesleyan Media Project’s dataset grew from 164 cases to more than 220.

The associated spending climbed from nearly $80 million to more than $137 million. Disclosure remained absent from roughly seven out of ten identified videos.

A meaningful improvement would require labels to rise faster than the total number of AI campaign ads. A stable ratio would show that adoption is outpacing transparency.

Researchers will also need to separate material fabrication from ordinary enhancement. That distinction can reveal whether the most deceptive uses receive the weakest disclosure.

The third signal is whether federal officials or industry groups establish a common standard after the election.

Campaign consultants have reasons to support consistency. Complying with different state definitions creates legal uncertainty for advertising distributed across multiple markets.

Federal action still faces constitutional limits and institutional constraints. Any rule must distinguish fraud from satire without becoming a general truth regulator.

A workable framework should begin with depictions of identifiable people performing actions that did not occur. It should also cover fabricated speech presented as authentic evidence.

Machine-readable provenance can support visible labels, but it cannot replace them. Metadata often disappears when content is recorded, cropped, or reposted.

Platforms can preserve credentials for participating creators. Determined campaigns can still distribute synthetic material through channels that do not maintain those records.

The 2026 numbers also should not be treated as proof that AI advertising changed an election result. Researchers have documented prevalence and spending, not a direct causal effect on votes.

Spencer Pratt’s Los Angeles campaign offers one cautionary example. His prolific AI content attracted attention, but he finished third in the primary.

Attention is not persuasion. A memorable synthetic clip can anger viewers, damage its sponsor, or disappear inside a crowded media environment.

However, uncertain effectiveness does not remove the transparency problem. Campaigns spent more than $137 million distributing these messages because they expected some political value.

Voters should examine the evidence inside an advertisement, not only its visual polish. A convincing image now carries less information about whether an event occurred.

Newsrooms should identify which exact elements were generated. Calling an entire advertisement “AI-made” can obscure the difference between editing and impersonation.

Campaigns should preserve source material and production records. Those records can support prompt corrections when opponents, regulators, or journalists challenge an advertisement.

The Wesleyan AI campaign ads count establishes a baseline for that accountability. It shows synthetic political media operating at funded, repeatable, national scale.

The open question is whether disclosure becomes a normal production requirement or remains a strategic choice. Watch the rulings, the final label ratio, and federal standard-setting.

Before sharing the next striking campaign video, ask three questions: Who paid for it, what was generated, and can its central scene be verified independently?