When voters went to the polls to elect Ireland’s next president, some of them may have been surprised to see Catherine Connolly’s name on the ballot.

Just days before, a deepfake video showing the eventual winner withdrawing from the race had circulated, imitating Connolly and multiple journalists within its fake reality.

In the Netherlands, two far-right members of parliament were found to be behind a Facebook page promoting deepfake images of their left-wing rival ahead of Sunday’s tight election, prompting apologies and recrimination.

This was the week that artificial intelligence hit two European electoral campaigns in a major way and exposed significant gaps in ongoing efforts to curb undue influence on voters.

There are concerns about what that means for European politics and for its voters, as politicians and regulators wake up to the arrival of AI-generated text and video content that has been part of U.S. political life for some time.

“The normalisation of such practices is worrying,” said Hannes Cools, assistant professor on the human factor in new technologies at the University of Amsterdam.

The Dutch election “is one of the first elections in Europe where we see that [the technology] has become an integral part in electoral campaigns in various ways,” said Claes de Vreese, a professor of artificial intelligence and society at the University of Amsterdam.

Shock jock

In a study of some 20,000 election-related posts in the Netherlands, researchers from the University of Amsterdam and the University of Mainz found that over 400 posts were AI-generated.

The party of far-right leader Geert Wilders, the Party for Freedom (PVV), came out on top in its use of AI. More than a quarter of the AI posts (120 in total) could be traced back to PVV-linked accounts.

Wilders kicked off the PVV’s campaign with an AI-generated video depicting a fictional future Netherlands living under Sharia law. Dutch weekly De Groene Amsterdammer reported that the video was made with OpenAI’s video generator Sora. 

When asked about the dominance of extremist or fringe parties in the use of AI, researcher Fabio Votta said, “There’s still a normative aspect of using AI.”

“For the far-right, a lot of their modus is norm-breaking and shocking. They don’t fear the reputation hit.”

Yet Wilders took the rare step on Monday of apologizing to Frans Timmermans, a former European Commission heavyweight and the leader of the GreenLabor-Left ticket, after it emerged through the Dutch press that two PVV members of parliament were behind a Facebook page spreading incendiary, AI-generated depictions of him. 

In one of the images, shown by Dutch daily De Volkskrant, Timmermans could be seen being led away by police in handcuffs. In another, he had his hands on a pile of money.

In Ireland, the fake video that saw Connolly announce her withdrawal from the presidential election was branded by the candidate as a “disgraceful attempt to mislead voters and undermine our democracy.”

Through a fake bulletin of Irish national broadcaster RTÉ, the video saw a deepfake version of Connolly declaring: “It is with great regret that I announce the withdrawal of my candidacy and the ending of my campaign,” with deepfake versions of two well-known TV presenters validating the news and discussing the impact.

Both Meta and Google-owned YouTube removed the Connolly video from their platforms without specifying how long it had been online. The Irish left-wing independent candidate won the election convincingly with 63 percent of the vote.

Depicting fictional events or attacking or discrediting other candidates are only two ways in which AI-generated content is being deployed to sway minds. 

Researchers also warn against a third, arguably more direct, method in which AI could influence election outcomes: users asking AI chatbots who to vote for.

With a large majority of voters typically undecided until the final days of the election, the Dutch data protection authority on Oct. 21 warned voters not to ask AI chatbots for voting advice, since these give a “highly distorted and polarized image of the Dutch political landscape.” 

“Chatbots are full of mistakes,” said de Vreese, adding that “they attribute various party positions to the wrong parties, and they also seem to have a kind of a suction effect” in a specific political direction.

An experiment showed that chatbots favored the GreenLeft-Labour ticket for voters on the left, while voters on the right were mainly directed to the far-right PVV.

 “People with a low literacy are particularly vulnerable to AI-generated disinformation,” said Cools.

Disclaimer

Regulators in Brussels have made election integrity, AI risks and online disinformation major priorities, a patchwork of ongoing efforts left them watching as the elections played out.

As the technology to generate AI content and the platforms to distribute them is mostly U.S.-based, all eyes are on Brussels for a bloc-wide response.

The EU’s powerful Digital Services Act puts some responsibility on platforms to tackle risks to elections, and both Meta and Google have recognized generative AI as a major risk factor — likely contributing to their decision to take down the Connolly video.

But the requirements are driven mainly by concerns about misinformation, rather than by efforts to regulate how European political parties use generative AI to spread their messages. 

Labeling is also a big part of the response, as required by a separate EU law specific to artificial intelligence. Researchers at the University of Amsterdam flagged that a majority of the posts they tracked for the Dutch election lacked an AI-labeling disclaimer. For those who did, it was the platform that added it, not the political parties.

More laws that could deal with the matter are on their way.

The European Commission is drafting guidance for so-called high-risk AI systems that can pose a risk to people’s fundamental rights, which will enter into force in August 2026 at the earliest. “These guidelines will include a section on AI systems intended to influence election outcomes or referendums,” said Commission spokesperson Thomas Regnier. 

Developers of the most complex AI models, such as OpenAI’s GPT or Google’s Gemini, have already had to comply with a series of obligations since August, including mitigating “systemic risks” to democratic processes. 

Next month, Brussels will unveil another proposal, meant to support EU countries in upholding the fairness and integrity of election campaigns against foreign manipulation and interference. That is not expected to contain any binding legal requirements.

Eliza Gkritsi contributed reporting.