Political advertising is becoming one of the biggest everyday uses of generative video, and the disclosure rules have not caught up. On October 16, 2026, a Techmeme summary of a New York Times analysis by Stuart A. Thompson said US candidates and PACs, mostly tied to Republicans, have spent more than $137 million on 220 AI-generated campaign ads in 2026, and that about 70% of those videos did not disclose AI use. The report adds that Republicans have used AI content more often than Democrats, in forms ranging from subtle enhancements to convincing deepfakes.
We could not open the Times article itself, so this post treats the $137 million, 220 and 70% figures as the Times' claims and checks them against the research and reporting we can read. The short version: the direction matches what academics and other outlets found, the totals are larger than the earlier public tallies, and the real story is how little the labels, and the laws behind them, are doing.
TL;DR: claimed versus checked
| Item | NYT (as summarised) | Other sources we could read |
|---|---|---|
| Spending on AI ads | More than $137M | About $80M on almost 170 unique ads (Wesleyan Media Project, early October) |
| Number of ads | 220 | 164 to about 170, depending on the date of the tally |
| Share undisclosed | About 70% | 69% unlabeled by ad count; 31% of ads disclosed, 22% of spending disclosed (Wesleyan) |
| Partisan split | Mostly Republican-tied | 80% of ads and 83% of spending from Republican candidates or pro-Republican groups (Wesleyan) |
| Federal law | Not covered in summary | None; about 31 states require labels |
The undisclosed share is the figure that agrees most closely across sources. The dollar totals differ, which is expected: the earlier tally was published before the final weeks of the campaign.
How does $137M compare with the Wesleyan numbers?
The most detailed public dataset is from the Wesleyan Media Project. Its researchers, Travis Ridout of Washington State, Erika Franklin Fowler of Wesleyan and Michael Franz of Bowdoin, wrote up their findings for The Conversation, republished by Fortune on October 6. They used media reports, student coders and AdImpact data and counted about $80 million on almost 170 unique ads. Republican candidates or pro-Republican groups were behind 80% of the ads and 83% of the spending.
The Times figure of $137 million on 220 ads is roughly 70% more in dollars and 30% more in ads than that tally. The gap could come from a later cutoff as more ads aired, a wider definition of what counts as AI use, or different spending data. Without the Times methodology we cannot say which. What we can say is that both analyses reach the same conclusions: the use is growing fast, it skews Republican, and most of it goes unlabeled.
How did we get here? A short timeline
The count has moved quickly through the year, which is part of why different outlets quote different totals. In March, CNN reported that Republicans released an AI deepfake of Texas Democrat James Talarico, one of the first widely discussed fabricated videos of the cycle. In May, the synthetic Peggy Flanagan video aired in Minnesota despite the state's ban. By September, Axios and NPR were both running pieces on how cheap generative video had become for campaigns, and on September 25 an Associated Press analysis, republished widely, was headlined that AI ads are popular in the midterms but state laws have not stopped a lack of disclosure. The Wesleyan tally followed in early October with $80 million, and the Times analysis on October 16 moved the number higher again.
Read as a series, the pattern is steady growth in volume with no matching growth in labeling. That is the practical reason to treat any single total as a snapshot rather than a final number: spending in the last three weeks before an election is typically heavy, so counts taken a few weeks apart should differ.
What do these ads actually look like?
AI use in campaigns is not one thing. The Wesleyan authors describe a range from realistic deepfakes of politicians, including Donald Trump and Alexandria Ocasio-Cortez, to subtle visual enhancements, and some ads that use AI to generate crowds or constituents.
Al Jazeera's October 9 report lists concrete cases:
- In New York's governor race, Republican Bruce Blakeman released ads showing Democrats Zohran Mamdani and Kathy Hochul together, though neither appeared in them. Blakeman said the ads are labeled as AI and called them satire. The AI label appears for only a few seconds at the start, and the New York Democratic Party filed a complaint with the state board of elections.
- Sherrod Brown, the Democrat in Ohio's Senate race, appeared as an overstaying party guest, and Texas Democrat James Talarico appeared in a blue dress from The Sound of Music.
- In a San Francisco House race, Democrat Scott Wiener released an AI chatbot that answered about 5,000 questions about opponent Connie Chan, mocking her accent and citizenship. It drew accusations of bigotry and Wiener withdrew it, saying it "missed the mark."
So this is not only a Republican practice, even if the numbers say they use it more. NYU's Scott Babwah Brennen calls the low-quality, obviously fake end of the spectrum "slopaganda": content that tells voters what to feel about a candidate's character rather than giving facts.
For how provenance checks work on a viral political fake, see our post on fake AI images and SynthID in the Youngstown rally case.
Why is disclosure so weak?
Glass box revealing gears, a picture of transparency that AI campaign ads often lack
There is no federal law regulating AI deepfake ads. About 31 states have passed laws requiring labels, but the Wesleyan researchers found they do not seem to change behaviour. Across 35 states they tracked, only 31% of ads, and 22% of spending, disclosed AI use. Ads in states without such laws disclosed 32% of the time, versus 29% in states with them. The authors caution the comparison is imperfect because some ads fall outside the laws' scope.
Minnesota shows the enforcement problem. A synthetic video of Democratic Senate candidate Peggy Flanagan aired in May 2026 despite the state's deepfake ban, and whether it broke the law is unclear. The wording of disclaimers also varies wildly, and as the Blakeman example shows, a label that flashes for a few seconds technically exists without informing anyone.
Public opinion is not the obstacle. The Wesleyan article cites polling in which 78% of registered voters favour banning AI content that makes deceptive claims about candidates. The authors speculate that the partisan gap in AI use may reflect differing philosophies, with Democrats generally favouring disclosure rules and Republicans a free-market approach, noting that in 2022 only Senate Democrats and two independents voted to advance the DISCLOSE Act. That is their interpretation, not a finding.
Experts quoted by Al Jazeera disagree on the fix. Ilana Beller of Public Citizen says a federal law is still needed and wants AI labels as clear as food labels. Adam Rose of Starling Lab argues AI ads should not be banned and that a candidate's judgment calls reflect on the candidate.
What does this mean for people who build or use AI?
Three things stand out for this audience.
Provenance tooling is the only scalable check. If labels are optional and unenforced, readers need other ways to test a clip. Content Credentials and watermarks help when the generator embeds them; our explainer on what C2PA Content Credentials are covers what they do and do not prove, and Google's public SynthID website guide shows how to check an image or clip. Our watermark removal coverage is a reminder that these marks can be stripped.
Labs are positioning on elections. Anthropic published its own work on Claude safeguards for the 2026 midterms and has started briefing 2028 candidates through its presidential engagement program. Campaign video is mostly made with other tools, which is why labels matter at the point of publication, not only at the model.
Likeness is the next legal front. Ads that put a real politician's face and voice into a scripted scene sit next to the celebrity cases we covered in likeness detection on Instagram and YouTube. Expect likeness and election-law claims to merge after November 3.
How to check a political video yourself
Label card with small bars, like the AI disclosure label a political video should carry
- Look for a disclaimer, and scrub the first and last seconds, since labels are often brief.
- Find the sponsor's own post or channel and read how it describes the clip.
- Check provenance tools if the generator supports them, and treat a missing label as no information.
- Search for the quote or scene in reputable reporting before sharing.
- Report suspected violations to the state elections board, as the New York Democratic Party did.
If you are a developer building tools that create or place political content, assume a disclosure rule will eventually reach you, and build labeling in by default. For teams running AI agents that publish on their behalf, AgentBeam is the agent security platform from the explainx.ai team that blocks risky agent actions before they execute.
What to watch next
The midterms are on November 3, 2026. Watch for the Times methodology, updated Wesleyan counts after the final weeks of spending, the outcome of New York's complaint over the Blakeman ads, and any federal movement. We will update this post if the Times figures are published in a form we can verify.
Figures reflect reporting available on October 16, 2026. The New York Times analysis was not accessible to us, and its numbers are attributed to a summary of it.
Related reading
- Fake AI images and SynthID in the Youngstown rally case
- What is C2PA Content Credentials?
- Google's SynthID website: verifying AI media
- Anthropic's midterm safeguards for Claude
- Anthropic's presidential engagement program
- Celebrity AI deepfakes and likeness detection
- AI watermark removal tools
- Sources: Wesleyan researchers via Fortune and Al Jazeera