Opening Artifact: September 28, 2023
On September 28, 2023, an AI-generated audio clip impersonating Michal Šimečka, then leader of Progressive Slovakia, and Monika Tódová, a journalist at Denník N, spread on social media two days before Slovakia’s general election. The clip depicted a fictitious phone conversation in which the two discussed rigging the election, including buying votes from the Roma community. Denník N denounced the clip as fake, and AFP’s fact-checking department identified several indications of AI manipulation. The Slovak prosecutor opened an investigation; the creator remains unknown. Tódová testified in March 2024.
That is the documented sequence. What follows is an audit of the machinery that let it happen — not a claim that anyone designed the moratorium as a weapon.
The Statutory Clock: What Slovakia’s Moratorium Actually Prohibits
Slovakia’s election-campaign law imposes a 48-hour silence period before election day. According to a TASR summary of the law published August 2, 2023, the moratorium before the September 2023 early parliamentary election began on September 28, and the campaign ended at midnight that day. The law prohibits publishing information about candidates in their favor or disfavor in radio, television, on-demand audiovisual media, periodicals, non-periodical publications, and agency reporting for 48 hours before election day and on election day until polling closes. It also bans publishing election polls and surveys in that window.
The same TASR summary notes that a longer two-week ban on publishing polls applied before the 2020 election, and that a 2021 amendment restored the 48-hour window. The penalty structure is specific: political entities face fines of €30,000 to €300,000 for violating the campaign ban; media face €1,000 to €10,000 for violating the ban on publishing information about candidates.
The statutory text is clear about who is bound: broadcasters, publishers, and agencies. It does not name platforms. That gap matters.
The Platform Clock: Meta’s Policy Scope in September 2023
At the time of the 2023 election, Meta’s manipulated-media policy covered faked video — where a person has been edited to say words they never said — but not audio. WIRED reported that because the post was audio, it exploited a loophole in Meta’s policy. The Coalition for Women in Journalism (CFWIJ) reported the same: Meta’s policies on manipulated media only covered video deepfakes, so the audio clip was allowed to circulate. A request to Meta to remove the clip remained unanswered, according to CFWIJ.
This is the second clock: the platform’s policy-coverage boundary. A statutory silence window constrains what domestic broadcasters and publishers can say. It does not constrain what a platform’s moderation system will action. If the platform’s rule enumerates only video, an audio-only fake falls outside the review trigger regardless of how fast fact-checkers work.
The Fact-Check Clock: Latency and the Limits of Speed
Veronika Hincová Frankovská, project manager at the fact-checking organization Demagog, told WIRED that her team had never previously had to fact-check an AI-generated post. They traced the recording to an anonymous Instagram account, called experts, and used an AI speech classifier made by Eleven Labs. After a few hours, they were ready to confirm they believed the recording had been altered. Their label went up on Facebook.
That is the third clock: fact-checker turnaround. A few hours is fast by normal standards. It is not fast enough when the statutory silence window is 48 hours and the platform’s policy does not cover the format.
The Distribution Picture: What the Jan Kuciak Investigative Centre Found
The Jan Kuciak Investigative Centre analyzed the spread of the clip and found it appeared in 20 Facebook page posts. Most dissemination occurred through personal profiles, making tracking difficult, according to CFWIJ. The creator of the manipulated audio is unknown.
This distribution pattern matters for audit design. Page-post counts are a weak signal when most spread is through personal profiles. Attribution is weak when the creator is unidentified. The tractable thing to measure is process timing: when the content appeared, when the statutory window opened, when the platform’s policy applied or did not, and when fact-checkers published.
What the Record Does Not Show
No retrieved source establishes that Slovakia’s moratorium was designed as an attack window, or that the clip’s timing was coordinated with the legal calendar or with any political actor. The defensible framing is narrower: the 48-hour window constrained legitimate rebuttal by domestic media and political actors, while the platform’s policy left the audio format uncovered. The collision of those two rules — not intent — is what the record supports.
Nor does the record show that the deepfake changed the election result. The election was close; Progressive Slovakia lost to SMER. But no retrieved source measures the clip’s effect on voters.
The EU Context: DSA and the Limits of the Record
WIRED described the Slovak election as one of the first consequential votes after the EU’s Digital Services Act was introduced in August 2023. The Act introduced rules meant to force platforms to be more proactive and transparent in moderating disinformation. Richard Kuchta, an analyst at Reset, told WIRED that the law put pressure on platforms to increase content-moderation or fact-checking capacity, and that Meta hired more fact-checkers for the Slovak election. Whether that was enough is a separate question.
The EU AI Act (Regulation 2024/1689) was not available in the retrieved sources for this audit. Claims about its transparency provisions and their interaction with national election-silence laws are therefore omitted rather than asserted from memory.
An Audit Checklist for the Next Election Window
For media-systems readers who need to apply this the same day, record three clocks for any election period:
- The legal silence window. Quote the statutory provision directly. Identify who is bound — broadcasters, publishers, agencies — and who is not. Note the penalty structure. In Slovakia’s case, the law binds domestic media, not platforms.
- The platform’s policy-coverage boundary. Read the manipulated-media policy as it stands, not as it is described in press coverage. Enumerate the media formats in scope. If the policy names video but not audio, an audio-only fake falls outside the trigger. Test the policy against the formats actually circulating.
- The fact-checker’s realistic turnaround. Ask the fact-checking partner how long confirmation takes for audio, video, and image formats separately. A few hours is fast; it is not instant. Map that latency against the silence window.
Then label the hypotheticals as hypotheticals. If a platform’s rule enumerates only video, an audio-only fake falls outside automated and manual review triggers regardless of how fast fact-checkers work. That is a design lesson, not a case finding.
Verification Habit: Read the Policy, Not the Coverage
The next time a deepfake story breaks during an election silence period, do three things before repeating the frame. First, pull the statutory text and identify who it binds. Second, pull the platform’s manipulated-media policy and enumerate the formats it covers. Third, ask the fact-checker how long confirmation took. The gap between those three clocks is the attack window — not because anyone designed it, but because the rules did not line up.
FAQ
Did Slovakia’s moratorium ban all election discussion?
The law prohibits publishing information about candidates in their favor or disfavor in broadcast, print, and agency reporting for 48 hours before election day and on election day until polling closes. It also bans publishing election polls in that window. It does not name social media platforms.
Why did the audio clip stay up on Facebook?
At the time, Meta’s manipulated-media policy covered faked video but not audio, according to WIRED and CFWIJ. The clip was audio-only, so it fell outside the policy’s scope.
Was the deepfake timed to exploit the moratorium?
No retrieved source establishes coordination between the clip’s timing and the legal calendar. The record shows the clip appeared during the 48-hour window and that the window constrained domestic rebuttal while the platform policy left the format uncovered.
Did the deepfake change the election result?
No retrieved source measures the clip’s effect on voters. The election was close, but the record does not support a causal claim.
What happened to the investigation?
The Slovak prosecutor opened an investigation, and Tódová testified in March 2024, according to CFWIJ. The retrieved sources do not document an outcome beyond that.