Since 29 January 2024, South Korean election law has banned campaign deepfakes in the 90 days before a vote. Two election cycles later, the numbers show a system that removes a great deal and penalises almost nothing: the first complaint came only on 29 May 2025. The real cost isn’t in fines — it’s in permanent monitoring infrastructure.
On 29 May 2025, days before South Korea’s presidential election, the National Election Commission (NEC) filed a complaint against three people, including operators of YouTube channels. The allegations concerned 35 posts showing a candidate’s image in prison uniform and 10 videos packaged as television news bulletins, fronted by an anchor generated entirely by artificial intelligence. It was the first complaint ever filed under the country’s election deepfake ban. The ban had been in force for sixteen months.
Those sixteen months are the subject of this piece, because they say something precise about what it actually costs to regulate synthetic content during a campaign. The technology is familiar: generative models trained on large volumes of images, video and voice recordings, able to produce a recognisable face or voice in a situation that never happened. On the other side, trying to catch them, sits a hybrid apparatus — automated detection tools, including one developed in-house by the Korean National Police Agency and deployed in 2024, plus human review, plus direct channels to platforms. That second infrastructure, not the ban itself, is the interesting line item.
What the law actually says
Article 82-8 of the Public Official Election Act was passed by the National Assembly in December 2023, promulgated on 28 December and became operative on 29 January 2024. It prohibits producing, editing, distributing, screening or publishing deepfakes — artificial audio, images and video «difficult to distinguish from reality» — for election campaign purposes, in the ninety days before a vote.
It is worth being precise about what falls outside, because foreign descriptions of this law tend to inflate it. It is not a blanket ban on AI-produced election content. Obviously stylised imagery, recognisable satire, and material not tied to campaigning are outside the perimeter. And outside the ninety-day window, campaign use of AI remains lawful — it must simply be clearly labelled as virtual or artificially generated content. Korean lawmakers did not switch AI off in campaigns. They decided that in the final three months, indistinguishable realism is itself unlawful, regardless of whether the content is true or false.
That choice has been criticised domestically as over-legislation. The ban applies even where the person depicted has consented, and even where the candidate produces the content about themselves. Commentators have flagged a chilling effect on entirely harmless uses. In academic commentary, Korea is grouped with Singapore as a jurisdiction that criminalises deepfakes per se during election periods, distinct from models that punish deception.
The numbers from two cycles
For the legislative elections of 10 April 2024, between 29 January and mid-February the NEC had already identified 129 AI items in violation. Across the full election period — 29 January to 10 April — the Commission cited 388 deepfake items flagged or subject to takedown requests.
For the presidential election of 3 June 2025, the scale changes. Between 4 April and 11 May 2025 the NEC sent 769 takedown requests to KISO and other recipients. The instructive detail is the tempo: as of 8 May the figure was 598; three days later, 769. A hundred and seventy-one in seventy-two hours.
Two serious caveats before anyone uses these figures. First, 388 and 769 are not comparable quantities — different time windows, different monitoring scopes, different elections. Second, and more important: these are takedown requests, not established violations. A takedown request is a fast administrative act; an alleged offence is something else entirely.
And this is where the structure fractures. After the 2024 legislative elections, the number of cases referred by the NEC to investigators under Article 82-8 was zero. Zero files were passed from police to prosecutors as well. The reason reported by the Korean press is the difficulty of the evidentiary requirements: you must simultaneously prove that the content is difficult to distinguish from reality and that it was made for campaign purposes. Two separate findings, both contestable, about material that often circulates anonymously. By 12 May 2025 police were investigating 8 videos involving 18 people. Then, on 29 May, the first complaint.
Where the money goes
The economics of this story are not where you would look for them. There is no reliable public data on compliance costs borne by platforms, on campaign spending for AI tools, or on any supposed impact on Korea’s AI industry: anyone quoting such figures is estimating. What is plainly visible is the apparatus the state had to build to make a ninety-day ban workable.
- A dedicated NEC advisory panel on deepfakes.
- Direct lines to platforms, including Naver and Google.
- Investigation centres stood up at 278 police stations.
- A proprietary detection tool built and deployed by the national police in 2024.
None of these switch on and off with the electoral calendar. An expert panel, an operational channel with a platform, a detection classifier are permanent assets: they must be maintained, updated and retrained as generative models improve. The ban lasts ninety days; the infrastructure that makes it enforceable runs all year. That is structural spending against an intermittent window — with a conversion rate into actual proceedings that, so far, is close to zero.
There is also a technical problem no budget solves. Automated deepfake detection is a statistical classifier: it produces probability scores, not verdicts, and its accuracy degrades when new generative models appear that were absent from its training data. A system that flags today’s fakes reliably can go blind against those of six months from now. For a fast takedown, a score may be enough. For a criminal prosecution, where the defendant can challenge the method, far less so. The gap between 769 takedowns and one complaint is not investigative laziness: it is the distance between two standards of proof.
The 2026 test
Korea’s AI Basic Act has been in force since January 2026. It is a general statute on artificial intelligence, not an anti-deepfake measure. One provision, Article 31(3), does however require notification and labelling of AI content that is difficult to distinguish from reality, with an administrative penalty of up to 30 million won — roughly 20,000 dollars — for failures of transparency.
The local elections of 3 June 2026 are the first round in which both frameworks apply together: a criminal ban inside ninety days, a labelling duty at all times. Two different logics — punish the production, or mandate the disclosure — coexisting on the same piece of content for the first time. Meanwhile, as of March 2026, an amendment is under discussion in Korea to extend the ban to elections for school superintendents, currently uncovered.
That last detail is the one worth keeping. A law written to defend a national vote left out a minor electoral contest, and two years later legislators are scrambling to close the gap. Not because anyone was careless, but because a ban built by perimeter — these elections yes, those no, within this window — ages at the speed at which access to generative models widens. The penalties, meanwhile, remain a theoretical deterrent: the statute provides for up to seven years’ imprisonment and fines of 10 to 50 million won, but across two election cycles nobody has yet seen them applied.