Survey identifies 554 deepfakes in 2026 election campaigns, most published by ordinary users

During the campaign for the first round of the 2026 general elections, at least 554 deepfake-type videos and images were published on social media, according to a survey by the VigIA project, carried out by the Lupa checking agency, which monitors the use of AI in the election.

Of these, 29 were disseminated through candidate or party profiles, which could lead to sanctions.

Deepfakes are content produced or edited with AI tools from the image of a real person, manipulated so that they appear speaking phrases or performing actions that are not real.

The Superior Electoral Court banned these publications, which could even lead to the revocation of the candidacy or mandate, but also noted that, to do so, they must be characterized as political propaganda and present verisimilitude capable of deceiving the voter.

Challenge

Report published this Thursday (1st) by the VigIA project identified that deepfakes were the main content created or edited with AI and published on social media during this electoral period.

They correspond to 60% of the total of 920 publications collected between August 16th and September 28th.

Among the most common deepfakes, videos in which journalists announce invented news and celebrities falsely express support for candidates stand out.

The majority of fake posts (525) were made by ordinary users, including many anonymous ones. For the researcher responsible for the report, Beatriz Farrugia, this data reveals a challenge for the application of electoral law:

"If we want to block the publication of deepfakes in any way due to the risk it poses to people's understanding of political events, we need to consider that most of this content comes from user accounts and anonymous avatars. Here the Court would need to work with the platforms to try to identify these people"

Signage

The report also identified another practice that contradicts electoral legislation. The use of AI is permitted as long as the parts created or manipulated with the help of these tools are flagged, which was not the case in most cases.

Of the 920 content identified, only 371 were properly labeled as material made with AI, around 40%. Considering the 549 irregular publications, 75 were made by candidates or parties.

VigIA in numbers

In total, VigIA collected 1,941 suspicious political-electoral content throughout the period analyzed. So far, 972 images and videos have been analyzed, with 920 confirmed cases of AI use.

The investigation is carried out in partnership with the Artificial Intelligence Laboratory (Recod.ai) at the State University of Campinas (Unicamp).

Three AI detection tools are applied to look for signs of use of these tools, and then the result is screened by researchers.

In addition to deepfakes, there is also synthetic content, in which technology was used to create or modify aspects of the publication, without attributing speech or actions to a person that did not happen.

For the founder and Intelligence and Training manager at Agência Lupa, Cristina Tardáguila, the numbers show that AI "is already part of everyday political communication, and not just isolated episodes"

"When we find almost one content with AI per hour and we see that the majority does not indicate the use of technology, we have a dimension of the challenge that this scenario represents not only for voters but also for supervisory bodies", he adds.

In general, the promotion of applications was the most frequent purpose among the content identified by VigIA, appearing in 414 of the 920 cases (45%). However, a quarter of the content contains attacks against candidates or other people.

There is also a similar proportion of pieces where AI was used to produce humor or satire content.

Elections

The national scenario was further explored, with a strong focus on the presidential elections. Accordingly, the current president and candidate for re-election, Luiz Inácio Lula da Silva (PT), was the main public figure portrayed in the pieces, appearing in 379 publications. Next comes Flávio Bolsonaro (PL), Lula's opponent, with 190 appearances.

In Lula's case, 39% of the publications were humorous or satire and another 34% sought to attack his candidacy. Among the publications with Flávio Bolsonaro, 40% sought to favor him and 28% contained humorous or satirical content.

The predominance of common users as main publishers remained in the general analysis, and 47% of authors have this profile.

However, the report also highlights that 27% of publications came from unknown profiles, a higher proportion than the 24.5% referring to the candidates' official accounts.