From 50,000 dollars to 187 million
On 3 November the United States votes in its midterm elections. The result will come later; here I look at how the campaign was run, and at a tool that in two years has gone from a curiosity to a line in the budget.
On 29 October the Wesleyan Media Project, the Wesleyan University centre that measures American political advertising, updated its count of ads made wholly or partly with artificial intelligence. There are 289, with more than 187 million dollars spent on them. Two weeks earlier there were 220, aired in 34 states, for more than 137 million. In the 2024 election cycle the same group had found less than 50,000 dollars' worth.1
The figure needs careful reading. Those 187 million dollars are the money paid to air ads in which AI produced at least one part, a face, a voice or a scene. And it is a floor, because the researchers only count ads that disclose AI, those reported by the press or by AdImpact, and those their coders identify with certainty. They say so themselves, calling the figure "almost certainly an undercount".
The growth is easy to see in the 2026 counts. In May the group had found 33 ads worth about 20 million dollars, on 4 September 164 worth almost 80 million.2
Spending on election ads made with AI, in millions of dollars
Who uses them, and for what
Use is not spread evenly between the parties. 82% of the ads and just over 90% of the spending, 169 million dollars, come from Republican candidates or groups, 13% of the ads from Democrats, the rest from independents and from groups on issues not tied to a party.1 Travis Ridout, one of the project's three directors, explains it this way: the Republican side views AI more favourably, while Democratic candidates and groups face a higher reputational risk.34 Gizmodo adds that many Democrats are calling for tighter rules on AI. The figure measures who adopted the tool first. It says nothing about the accuracy of individual ads, and the researchers themselves point out that AI is also used for harmless things, such as stock footage, and saves filming costs for candidates with small budgets.
A significant share, though, is used to make opponents say things they never said. According to the same outlet, citing the researchers, nearly half of the ads identified contain deepfakes of real politicians, often national figures placed alongside local candidates.4 In Texas an ad from the campaign of Ken Paxton, the Republican Senate candidate, shows Democrat James Talarico, recreated with AI, delivering an invented monologue.3 Talarico responded by ending one of his own ads with the line "Made by humans. Not AI."4 In Pennsylvania the Republican candidate for governor, Stacy Garrity, had Democratic Governor Josh Shapiro praise data centres, while stating that the video was generated.3
Disclosures like Garrity's are the exception. In mid-October only 27% of the ads carried a disclaimer.1 In September the ads without one were 69% of the total and took 78% of the spending.2 On average, then, the ads that disclose nothing also have more money behind them.
In the United States, each state decides
There is no federal law requiring AI to be disclosed in political ads, and the Wesleyan Media Project points this out in every report.2 The rules are those of individual states. According to Public Citizen's tracker, updated on 21 October, 33 of the 50 states have passed at least one law on election deepfakes, and in Hawaii a court has already struck one down.5
The New York case shows where the rule jams. The state requires deceptive AI content in political ads to be disclosed. According to Gizmodo, Bruce Blakeman, the Republican candidate for governor, has posted at least 22 AI-made ads on social media since April. In August the state Democratic Party filed a complaint with the elections board, and Blakeman's campaign replies that satire needs no disclaimer.4 Whoever is right, a complaint filed in August has to be decided before a vote set for 3 November, and meanwhile the ads keep circulating.
In Europe and in Italy the rules are stricter
Europe starts from a different place. Since 10 October 2025 the EU regulation on the transparency of political advertising has applied: every ad, online and offline, must be labelled and say who pays for it, what it costs and, where targeting is used, which audience it is aimed at. On 12 October 2026 the Commission opened a pilot version of the European repository of online political ads.6
On AI, Article 50 of the AI Act applies, which I wrote about in the piece on SynthID and watermarks. Since 2 August 2026 anyone who publishes a deepfake must disclose that it is generated or manipulated.7 For "evidently" satirical or artistic works the obligation is lighter but does not disappear: it is enough to disclose it in a way that does not spoil the work.8 It is the European answer to Blakeman's defence, and it seems the right one to me.
On Italy's national TV channels the American-style attack ad does not exist. Since 2000 the par condicio law has allowed national TV and radio stations to air only free, self-produced political messages, one to three minutes long, in dedicated slots and marked "self-produced message" with the name of whoever commissioned it.9 AGCOM's AI Committee, in its June report, notes however that the law was designed for radio and television and covers neither social media nor the way algorithms decide what we see.10
Italy's AI law, Law 132 of 2025, added two criminal tools. Anyone who uses deception to stop someone exercising a political right, such as voting, or to push them to exercise it against their will already faced one to five years: if the deception goes through AI, the sentence rises to two to six. And anyone who spreads, without consent, images, video or voices faked with AI that cause unjust harm faces one to five years.11 I wrote about how oversight is split between authorities in Who will police AI in Italy.
The rules on political ads made with AI
| Where | What it provides | The weak point |
|---|---|---|
| United States | No federal rule; election deepfake laws in 33 of 50 states | Disclaimers are rare and complaints are slower than the campaign |
| European Union | A label and information on who pays for every political ad; deepfakes to be disclosed from 2 August 2026 | For satire a discreet notice is enough; the obligation is only a few months old |
| Italy | Only free, signed messages on national TV; two to six years for deceiving voters with AI | The 2000 par condicio law does not cover social media and algorithms |
The problem is time
So the rules exist. The problem is enforcing them before the damage is done.
In Slovakia's 2023 election a fake audio clip of pro-European leader Michal Šimečka, in which he appeared to discuss vote rigging, came out two days before the vote, during the electoral silence period, and spread mainly on Telegram.12 Šimečka lost. The scholars who reconstructed the case consider it reductive to say that a deepfake decided the vote, and point to deeper causes such as distrust of institutions. But the episode points to the weak spot of any rule. The fake arrives when the rebuttal no longer has time to reach voters. According to AGCOM's AI Committee a rebuttal takes 18 hours at best, and during the 2024 European elections AGCOM had already received several reports of possible deepfakes.10
Italy's electoral silence rule dates from 1956 and bans rallies, campaign meetings and new posters on the day before the vote and on polling days. Campaigns, however, moved onto phones long ago. At the constitutional referendum of 22 and 23 March 2026 television was again the first source of information, but almost half of adults also used the internet, and a fifth used social media and video platforms. Among algorithmic sources AGCOM now also counts chatbots.13 I wrote about this in the piece on news that comes from chatbots.
How Italians followed the March 2026 referendum
Some urge caution about overstating this. Sayash Kapoor and Arvind Narayanan of Princeton examined 78 cases of AI use in the 2024 elections worldwide: half had no deceptive intent, and the outright fakes could have been made without AI for a few hundred dollars. In their view the bottleneck for misinformation lies in distribution more than in production.14 It is a serious objection, and the 2026 American data put it to the test, because with 187 million dollars distribution can be bought.
Copying a face and a voice now costs almost nothing. What remains is deciding whom to trust, and that choice is made in the last days of a campaign, when there is least time to check.
What I think
I am not taking sides between parties. What worries me is a figure that concerns everyone. Left to the goodwill of campaigners, the disclaimer appears once in four. A voluntary label, in a race where every vote counts, does not hold. It has to be an obligation, apply to satire too and carry sanctions that arrive before the vote, not months later.
On this Europe is in better shape than the United States, because the obligations exist and apply across the Union. In Italy the paid TV ad is not the main risk, and par condicio has kept it in check for more than twenty-five years. The risk is the short video circulating on social media and in chats in the last 48 hours, with no visible sponsor. This is exactly what AGCOM's AI Committee has in mind when it calls for a protocol with the platforms to take down political deepfakes within two hours of a report, a team working around the clock during the campaign with the Postal Police and the platforms, and common standards for labels, before the 2027 general election, to avoid what it calls "an Italian Slovakia case".10 I agree. Two hours is little time to decide what is fake, and someone will have to answer for every removal, but 18 hours is too long.
I would add one thing that needs no new law. Before the next campaign the parties, all of them, could commit publicly to disclosing every piece of content made with AI, satire included, and to keeping the originals online for anyone who wants to check.
- PartiesAlways discloseA clear notice on every piece of content made with AI, satire included, and the originals available for checking.
- PlatformsRespond within hoursA dedicated channel during campaigns, with takedown times for political deepfakes measured and made public.
- AuthoritiesA protocol ready in advanceAgreements signed before the campaign starts and someone who answers even the night before the vote.
- VotersBe wary of the last minuteA sensational video released just before the vote deserves an extra check: look for the original source before sharing it.
Italy's next general election is expected in 2027, and the campaign will start months earlier. The protocol with the platforms, and a number to report a deepfake to on the night before the vote, need to be prepared now, while no campaign is under way.