Original article (in Croatian) was published on 1/23/2025; Author: Ivica Kristović
Despite advances in AI-generated fake audio and video technology, Croatian politics was fortunately spared in this super-election year.
In the midst of the first round of presidential elections, at a press conference, while explaining what is true regarding the meeting between her husband Mile Kekin and Nikica Jelavić, the Možemo! platform presidential candidate Ivana Kekin played a very convincing audio AI-generated recording of Prime Minister Andrej Plenković expressing his support for her.
“What we are going to play for you now (…) was done in ten minutes, with zero funds, with voice-generating tool available on the Internet,” Kekin said at the press conference, before Prime Minister Plenković’s completely fake statement was played. Of course, Plenković did not support Kekin for president because HDZ’s candidate in the presidential elections was Dragan Primorac.
This recording clearly demonstrated both the power of this technology and the dangers it poses to democratic and electoral processes. Fortunately, nothing similar was used in this presidential campaign.
According to the data obtained by Faktograf from the Croatian Regulatory Authority for Network Industries (HAKOM), there were no recorded reports related to the use of AI content generation tools for the purpose of disinforming the public.
Slovak case
The night before election day, during the election silence, an audio recording of a speech made by a candidate began spreading across social networks and chat groups with a large number of followers, which deeply compromised the chances of that candidate. Because the elections were the following day, election silence was in effect, and the technology for detecting fake recordings had still not been sufficiently developed and reliable, so it was impossible to determine with 100% certainty whether the recording was authentic. Voters, motivated by the fact that it was not clear whether the recording was authentic, did not vote for that candidate just in case, and this prevented the candidate from entering the second round of elections…
This is a drastic case of election interference, enabled by technology that allows for the rapid creation and widespread distribution of fake recordings.
The first drastic case of using AI technology to generate fake recordings for the purpose of interfering in the election process occurred in September 2023 in Slovakia. There, two days before going to the polls, a fake AI-generated audio recording of a conversation between Progressive Slovakia party leader Michal Šimečka and journalist Monika Tódova appeared. In the fake recording, the two say the election will be rigged in favour of Progressive Slovakia.
Based on the statements of experts, fact-checkers from the Slovak AFP concluded that the recording was not an authentic conversation between the politician and the journalist. But immediately after the original announcement, this two-minute low-quality audio recording spread to thousands of Slovak social media profiles. According to pre-election polls, the Progressive Party was right behind Robert Fico’s SMER-SSD party, who promised to stop military aid to Ukraine. Before the elections, as many as 22 percent of voters were still undecided. In the end, the SMER won with 23 percent, and the PS was second with 18 percent, so Fico created a ruling coalition and became prime minister again.
As for the fake recording that was shared on social networks before the elections, the Slovak judiciary has not yet determined who released it to the public. The criminal complaint was filed immediately after its publication by journalist Monika Tódová, but the police rejected it for the first time after a few weeks, concluding that, although investigators found it to be a clear fake, it was not convincing and therefore could not affect a significant number of voters. However, the prosecutors had a different opinion and ordered the police to reopen the case. After a year-long re-investigation, the police concluded that they could not find the source and initiate criminal charges against a particular person. Such a conclusion has been appealed by journalist Tódová, and the prosecutor needs to decide on her appeal again.
The pitfalls of previous Croatian elections
While the focus of our text is on fake AI-generated recordings, it’s worth recalling the cases of secretly recorded tapes that have previously surfaced in Croatian politics, which political analysts have determined had varying degrees of impact on election results. At the time, these were authentic, but secretly recorded recordings, which shows that such content can affect the election process and the decisions of voters at the polls.
A month before the early parliamentary elections in 2016, scheduled due to the dissolution of the short-lived government of non-partisan Prime Minister Tihomir Orešković, who was appointed by the HDZ Tomislav Karamarko and the Most Bože Petrov after months of negotiations and after very tight election results, a secretly recorded recording of the conversation of then SDP President Zoran Milanovic with veterans’ associations was released to the public, at which the candidate for the new prime minister candidate insulted neighbouring countries.
The secret recording of the nominally left candidate with veterans’ associations, which are known as bastions of right-wing politics, which only a few months earlier “camped” in front of the Ministry of Veterans’ Affairs and asked for the dismissal of Milanović’s minister and his associates, had attracted great public attention and probably influenced the results of the September elections which Milanović lost to the president of HDZ, Andrej Plenković. Then, in 2016, Milanović confirmed the authenticity of the recording of the conversation with the war veterans.
A few days before the parliamentary elections in April 2024, the weekly Nacional published on its website, and in print the day before the election silence, a secretly recorded recording of the speech of HDZ president Andrej Plenković from a closed party meeting at which he insulted the president of Domovinski pokret, Ivan Penava, and other political figures. As in the case of Milanović, Plenković did not dispute the authenticity of the recording.
The function of bots
In the recent elections, there were no AI recordings, but the use of AI tools was noticed. In the first round of presidential elections, the social network TikTok confirmed to Faktograf that it had removed over 7,000 bots, i.e. inauthentic followers, from the profile of Dragan Primorac, the independent candidate endorsed by HDZ and their partners.
On the eve of the second round of presidential elections, Primorac accused Milanović of working for foreign interests on the basis of research published by the London Centre for Information Resilience (CIR), which stated that “Croatian researchers, in partnership with the CIR, have discovered indications of a pro-Kremlin disinformation campaign that uses a large network of bots before the second round of elections in Croatia.”
“Pro-Russian, anti-EU, and anti-NATO bot networks were found to consistently praise President Zoran Milanovic and amplify pro-Milanović content while disseminating anti-EU and anti-NATO messaging,” said the report, which prompted Primorac to say that even “the results of the presidential elections are endangered.”
However, the whole report turned out to be quite unreliable, and some accounts that the report said were bots were found to have real people behind them. Also, social networks X, TikTok and Meta confirmed that they found no evidence of inauthentic accounts or bots that would affect the Croatian elections, and the only bots that were identified were those on Dragan Primorac’s TikTok profile.
HAKOM: The election did not record the use of AI content for the purpose of spreading disinformation
When it comes to creating videos with false statements by politicians and other public figures, which are easily spread on social networks, Faktograf has detected dozens of them in the last year and a half.
These recordings, however, did not aim to interfere with political processes; rather, they were used as “evidence of the authenticity” of various frauds – ranging from the sale of suspicious drugs to the promotion of non-existent investment platforms. In the recent parliamentary, European and presidential elections, there were no such drastic examples of the use of generated AI recordings in the electoral process. This was confirmed by the Croatian digital services coordinator HAKOM (Croatian Regulatory Authority for Network Industries), whose tasks include the implementation of the EU’s Digital Services Act (DSA).
“We did not have any recorded reports (either directly or indirectly) related to the use of AI content generation tools for the purpose of disinforming the public”, said Lidija Antonić, Digital Services Regulation Coordinator at HAKOM.
Digital Services Act
As explained in the text on the Faktograf Association’s website, the Digital Services Act (DSA) was adopted by the European Parliament in October 2022 with the aim of preventing illegal and harmful activities on the Internet and the spread of disinformation. Among the risks that threaten social networks, the act also lists “actual or foreseeable negative effects on democratic processes, civil discourse and electoral processes, as well as public security”.
According to the Act, “providers of very large online platforms and very large online search engines should with due diligence identify, analyse and assess all systemic risks in the Union arising from the design or functioning of their service and its related systems, including algorithmic systems, and from the use of their services”, and when conducting risk assessments, they must also include “all actual or foreseeable negative effects on civil discourse and electoral processes and on public security”.
As we stated earlier, the appearance of such generated AI recordings with falsified statements of election participants and their dissemination to a large number of social network users would certainly have negative effects on the election process, which is mentioned in the Act.
Croatia is late with implementation
Croatia is significantly late with the implementation of this Act. The procedure was initiated at the end of the mandate of the last Government of Andrej Plenković, and it was only in January 2024 that the Ministry of Economy published a public consultation on the Draft Act on the Implementation of the Digital Services Act. There were no other developments until July, so the European Commission initiated proceedings against Croatia for violating EU law. At the end of August, the new Government, in which the Ministry of Justice, Administration and Digital Transformation took over the care of the DSA, adopted the Proposal of the Act, and at the end of September the Croatian Parliament discussed it in the first reading.
According to that Proposal, HAKOM is designated as a national Digital Services Coordinator that meets strict conditions (independent of any external influence, autonomous in managing its budget, possesses sufficient technical, financial and human resources).
The Croatian Parliament has not yet adopted the Act on the Implementation of the Digital Services Act, nor has the Government sent it a Final Proposal after the Proposal from August 2024. However, even during the presidential elections, HAKOM was able to perform its role of the digital services coordinator based on the decision made by the Government which appointed it on 14 February 2024.
HAKOM: The platforms acted in accordance with their obligations to the DSA
In addition to informing us that they received no reports related to the use of AI content generation tools to spread disinformation during the elections, HAKOM stated that ‘the platforms acted in accordance with their obligations under the DSA”.
In response, HAKOM also explained to us what they were doing in its role as the DSA coordinator to ensure that there was no interference with the electoral process.
“Since the pre-election period, HAKOM, in its role as Digital Services Coordinator, has maintained regular communication with the European Commission and organized a preparatory round table with very large online platforms (VLOPs), state bodies and relevant civil society organizations (including Faktograf) with the aim of ensuring the integrity of the electoral process and establishing communication channels to ensure a timely response in case of need,” they said in the response.
“Presidential elections without significant incidents”
They also confirmed that the first and second rounds of elections went without significant incidents. “The platforms responded in a timely manner to the limited number of reports received, including four reports from GONG, one report from the State Election Commission (DIP) and one report from the recipient of the service. Reports were generally correctly addressed directly to the platforms, while HAKOM was informed as the coordinator. In addition, DIP made certain recommendations to the Meta and Google platforms in order to make it easier for service recipients to access verified sources of information related to electoral issues and inform them of the period of electoral silence”, HAKOM stated in its response.
We also asked them whether there were any systemic risks recorded in the elections on social platforms. “In relation to the query related to systemic risks, we clarify that the analysis and assessment of the same is not up to the coordinators of the Member States, but exclusively on the providers of very large online platforms and very large online search engines. In the context of Croatian elections, none of the platforms recognized that there would be systemic risks arising from the design or functioning of their service and its related systems, including algorithmic systems, and from the use of their services”, HAKOM responded. In the end, they also confirmed to us that they found no evidence of AI-generated recordings being used to spread disinformation during the election process.
Although the elections held in 2024 and early 2025 in Croatia did not record the use of AI-generated recordings used for spreading lies during the election process, the Slovak case of fake recordings of politicians and journalists on the eve of the elections best illustrated why states need to prepare for such scenarios and strengthen institutions that fight such harmful practices. This technology is already used in Croatia: politicians have been put on fake recordings, but still only for the purposes of selling fake products, not yet in the field of political fights.
In May, Croatia will face another test of its resilience against this harmful practice: the upcoming local elections. These are often more important to political actors than the presidential ones because the number of participants in the political match is much higher, and thus there is a larger number of politicians who could conclude at some point that “the end justifies the means” and turn to this new technology in an attempt to win a mandate.