There is a point where two of the biggest technology stories cross — the regulation of artificial intelligence and the integrity of elections — and it is one of the most delicate fronts for democracy: the use of generative AI to fabricate false electoral content. Here the gap between technology and law takes a particularly sharp form, because it is not only that regulation arrives late, but that the very tools democracies used to protect their elections — campaign silence periods, broadcast rules — were designed for a media ecosystem that no longer exists. With 2026 densely packed with elections across continents, that mismatch is no longer theoretical.
Begin with the scale of the phenomenon, because the numbers are stark. Deepfakes have crossed a threshold in 2026: they have shed the tell-tale glitches that once gave them away — the unnatural shadows, the off-kilter light, the wrong number of fingers — and are now accessible to anyone with a smartphone. A cybersecurity firm documented dozens of pieces of AI-generated deepfake content targeting public figures across nearly 40 countries in a single year, disproportionately focused on elections. Synthetic content — any image, audio, or video created or altered with AI to appear real — has stopped being a hypothesis and become campaign material.
The cases that already happened
There is no need to speculate about the future; the recent electoral cycle already has a record. In Ireland’s 2025 presidential election, a deepfake video falsely showed the eventual winner withdrawing his candidacy, complete with fabricated footage of broadcasters “confirming” the news — released days before the vote. The Netherlands saw roughly 400 AI-generated synthetic images used to attack political rivals. In Brazil’s October 2024 municipal elections, monitors at the DFRLab identified 78 instances of confirmed or alleged synthetic content targeting local candidates, spread across Facebook, Instagram, X, TikTok, YouTube, and WhatsApp. After the first round of Poland’s 2025 presidential election, AI-generated images featured in viral videos alleging fraud. In the United States, automated calls with a cloned voice of a national figure had earlier urged voters to skip a primary.
The pattern analysts describe is revealing, because it is not random. Synthetic electoral content tends toward a short, emotional format — clips of a few seconds, built for mobile and for chain-sharing — with surgical timing: it is released on the eve of the vote, when the capacity to respond is lowest, and amplified in targeted fashion through a constellation of anonymous accounts and influencers. The design is deliberate: maximize damage at the moment of least defense.
The regulatory patchwork
Faced with this, jurisdictions have reacted — but in strikingly different ways, producing a global patchwork rather than a common answer. The European Union took the transparency route: under Article 50 of the AI Act, enforceable from August 2026, AI-generated or manipulated content must be labeled and synthetic interactions disclosed, with fines reaching up to 6% of global revenue; the Digital Services Act separately obliges platforms to mitigate AI-driven disinformation during election cycles. Brazil took the electoral-law route: its Superior Electoral Court adopted a resolution for the 2024 municipal elections that banned deceptive deepfakes, required AI-generated propaganda to be labeled, and treated misuse as potential grounds for annulling a candidacy. France allows courts to remove misleading AI-generated political content; Singapore mandates labeling and removal; China requires providers to label and remove deceptive synthetic media.
The United States is the cautionary case. Federal free-speech protections complicate deepfake regulation, so the action moved to the states — around twenty now regulate campaign-related synthetic media, with different definitions, timeframes, and penalties. The inconsistency is itself a problem: content legal in one state may be a felony in the next. And the constitutional limits are real: California’s AB 2839, which barred election content that “seriously misleads voters” and required even satire to carry disclaimers, was struck down by a federal judge as unconstitutional. The result is a live compliance headache for anyone operating across borders, and an uneven shield for voters.
The deeper gap: an obsolete defense
Here is the point worth underscoring, because it goes beyond any specific rule. The structural problem is not only that rules are missing, but that the classic tools of election protection were built for another world. Take the campaign-silence period — the quiet days before a vote when campaigning is barred, common across many countries. That mechanism was designed for a mass-media ecosystem: switching off the campaign meant pulling TV ads, taking down posters, ending the last rally. It made sense when political communication flowed through identifiable, controllable channels.
None of that stops a generative-AI model today. A system can produce hundreds of versions of a message and push them into circulation in minutes — anonymously, in targeted fashion, precisely during the silence period, when the law forbids responding and institutional reaction is slowest. The response, when it comes, is analog against a digital problem: criminal complaints, court orders to take posts down, rebuttals — all at a speed incompatible with virality. By the time a falsehood is debunked, it has already done its work.
Two readings, with comparable weight
The debate over how to respond admits two legitimate positions, worth presenting without tilting the scale.
Those who favor firm regulation — electoral courts, bodies like UNESCO, many lawmakers — argue that electoral integrity is a good worth strict rules: without a shared reality grounded in facts, they contend, elections cannot do their job. For them, banning deceptive deepfakes, requiring labels, and holding platforms responsible are measures proportionate to the threat.
Those who warn against overregulation point to real problems: the technical difficulty of detecting deepfakes — if experts struggle, how will an electoral judge in hours? — the risk that broad rules end up censoring satire, parody, or legitimate content (as a US court found with California’s law), and the danger of selective enforcement. There is also a perverse effect researchers already observe: as the public learns deepfakes exist, people begin to doubt everything, including the authentic — what some call the “liar’s dividend,” whereby a politician can dismiss a real recording as fake.
It is not for this outlet to decree the exact balance. What can be stated is that both positions describe a real dilemma: protecting electoral truth without turning the referee into a censor, and without the mere existence of the technology eroding trust in all evidence.
What this reveals heading into a year of elections
For a year as dense with elections as 2026, this is a standing concern, not a footnote. The contests ahead will be among the first held with generative AI models fully accessible, cheap, and sophisticated. The question is not whether synthetic content will appear — it will — but whether institutions will have adapted their defenses in time, or keep responding with tools from another era.
The verifiable fact is that electoral synthetic content has grown fast, that it has already been used in concrete contests across continents, and that the regulatory response — from the EU’s labeling rules to Brazil’s bans to a fractured US patchwork — exists but collides with a structural problem: classic electoral defenses were not built for this speed or this anonymity. Whether democracies close that gap will depend on decisions not yet fully made: on whether electoral referees acquire real technical capacity, on whether platform responsibility becomes effective, and on whether voters’ digital literacy keeps pace with the threat. As in every story of this kind, what is decisive is not the existence of the deepfake — which is measured and rising — but whether the institutional response learns to move at the speed of the problem, instead of arriving, as it has so far, once the damage is done.