AI propaganda: a lot of noise, but how many votes?

$137 million: that is the sum spent to air 220 election ads made at least partially with AI in the United States, according to the Wesleyan Media Project [1], which examined campaigns up to October 11. Only 27% of them explicitly acknowledge the use of this technology. The sample includes 81% of ads coming from Republican candidates or organizations, which account for 88% of spending.

These $137 million do not represent the cost of producing the deepfakes, but the cost of broadcasting them. A considerable investment that says nothing, however, about their electoral effectiveness.

An Axios investigation of October 19 [2] illustrates the phenomenon of AI-generated ads, often used to produce fakes: a Republican committee thus shows Democratic senator Jon Ossoff receiving cosmetic treatments. In Texas, an ad from Ken Paxton’s campaign shows his opponent James Talarico desecrating a Bible and uttering invented remarks. In Pennsylvania, a digital avatar of Governor Josh Shapiro attributes false statements to him.

Covert propaganda, for its part, no longer even needs to pay to display impressive results. On October 16, Bloomberg [3] documented a campaign of AI-fabricated videos, notably fake appearances by Robert De Niro, Julia Roberts or Cher against American Democrats. Researchers link this content to Operation Matryoshka, an influence network associated with Russian actors, without demonstrating state direction for these posts, however.

When robots also manufacture the audience

The most revealing aspect concerns audiences. One post recorded more than 50,000 views and a single repost within twenty minutes, before practically ceasing to grow, a recurring case. For researchers, these anomalies point to artificial amplification by automated accounts.

In other words, bots are not only used to spread propaganda, but also to manufacture the traffic supposed to prove its success. It remains to be seen how many real users saw this content, how many believed it and how many changed their minds, three very different scales.

A review of 25 empirical studies, published in May in Computers in Human Behavior Reports [4], provides a first counterpoint: political deepfakes would not be inherently more persuasive than other forms of disinformation. They can, however, cause confusion, distrust of information and generalized suspicion, including toward authentic content. Indeed, the “liar’s dividend” is probably one of the most immediately exploitable effects of these deepfakes: their proliferation makes it possible to contest authentic content, by claiming it was fabricated by AI.

The Computers in Human Behavior meta-analysis does not, however, mean that AI would be ineffective at persuading. A study published in PNAS [5], involving 14,779 British participants, shows that conversations with an LLM increase the probability of signing a real petition by 12.8 to 19.7 points, depending on the experiment.

Changing one’s mind is not taking action

More troubling still: according to the researchers, the methods that change opinions the most are not necessarily those that trigger the most action. Providing convincing information is therefore not necessarily enough to trigger action.

Another experiment, published in PNAS Nexus [6], reveals that warning a user that ChatGPT would be hostile to their political camp reduces its persuasive power by 28%, without making it disappear altogether: trust in the intermediary therefore matters too.

These results shed light on the mechanisms analyzed in our investigation of LLMs as a cognitive attack surface [7]. A falsified video broadcast on a large scale and a personalized conversation with an assistant mobilize neither the same drivers nor the same means of measurement.

AI makes it possible to fabricate lies, bots to simulate their popularity and assistants to prompt measurable behaviors. It is still necessary to distinguish between production, distribution, exposure and persuasion. One can buy ads, fabricate views and simulate popular enthusiasm. Buying convictions is another matter.

Sources and references

[1] GOP Lead in Senate Spending Narrows; Dems Catch Up in Senate Broadcast Airings, Wesleyan Media Project, October 15, 2026.

[2] The AI attack-ad nightmare is coming true, Axios, October 19, 2026.
https://www.axios.com/2026/10/19/ai-attack-ads-campaigns-elections-deepfakes

[3] Russia Is Using an Army of Bots to Push Fake US Election Videos, Bloomberg, October 16, 2026.
https://news.bloomberglaw.com/artificial-intelligence/russia-is-using-an-army-of-bots-to-push-fake-us-election-videos

[4] Exploring how political deepfakes trigger cognitive processes and downstream psychological and behavioral outcomes: A systematic review, Computers in Human Behavior Reports, May 2026.
https://doi.org/10.1016/j.chbr.2026.101066

[5] Artificial intelligence can persuade people to take political actions, Proceedings of the National Academy of Sciences (PNAS), October 20, 2026.
https://doi.org/10.1073/pnas.2611368123

[6] Perceived political bias in large language models reduces persuasive abilities, PNAS Nexus, October 9, 2026.
https://doi.org/10.1093/pnasnexus/pgag324

[7] Manipulation des LLM : comment l’IA devient une surface d’attaque cognitive (LLM manipulation: how AI becomes a cognitive attack surface), InCyber, October 2026.