When artificial intelligence stops being a campaign tool and starts deciding what voters see

On 19 February 2026, seven weeks before Hungary voted, Fidesz published a campaign video that no camera crew had filmed. Black and white, scored for strings, it told a story lasting under a minute: a father is mobilised, taken prisoner and shot by enemy soldiers while holding a photograph of his daughter. None of it had happened. None of it, in the ordinary sense, was a lie either. It asserted no fact a commission could strike out, named no one, invented no quotation. It showed a voter a future and let the fear do the rest.

Zsuzsa Detrekoi, writing in Frontiers in Communication in July, put the difficulty plainly. Existing legal frameworks, she argues, "are designed to address clearly defined harms such as incitement, deception, or voter interference," and so "they remain largely unable to respond to emotionally manipulative yet formally lawful political communication." The apparatus is built to catch falsehood. The Fidesz video was not false. It was fiction, and fiction has always been legal in politics.

Manufactured emotion in campaigns is old; Willi Münzenberg was manufacturing it in the 1920s. What has changed is the economics. A poster had to be designed, printed and pasted; a party broadcast was made once and shown to everyone. Because production was costly, the message was singular, aimed at an average voter who did not exist. Generative systems collapse that cost to roughly nothing. The same appeal can be recut into ten thousand versions overnight, each tuned to a segment, and the performance data returns fast enough that the system learns which framing moves which person before the campaign has finished its morning meeting. That is the shift worth arguing about, and it has little to do with deepfakes.

Deepfakes are the smallest part of it#

The fixation on synthetic video is understandable and misplaced. Fabricated footage is the most visible use of these systems and the easiest to legislate against, which is why it is the least interesting. The consequential work happens upstream of anything a voter would recognise as fake: ad targeting, turnout modelling, message generation at volume, bot networks, voter segmentation, campaign automation, and the quieter business of deciding what appears at all.

Consider a finding from the Dutch data protection authority, published on 12 March 2026, six days before the Netherlands held municipal elections. The AP tested four chatbots — ChatGPT, Gemini, Mistral and Grok — against established voting-advice tools such as StemWijzer and Kieskompas. Fewer than one per cent of the chatbots' voting recommendations mentioned a specific local party. Local parties take roughly thirty per cent of the vote in Dutch municipal elections. The regulator's conclusion was blunt: AI chatbots are not suitable for giving voting advice.

Nobody wrote a lie. No deepfake was produced. A category of political actor commanding a third of the electorate was simply, statistically, not there. The cause is dull and structural: national parties generate more text online, local parties publish less and update it late, and a model trained on the accumulated internet reproduces the imbalance as though it were the shape of the world. Dullness is not innocence. If more voters enter politics through a chatbot, the model's training distribution becomes a de facto electoral threshold.

Full Fact's 2026 report describes the commercial response forming around this. "A new commercial ecosystem has emerged," it notes, "in which brands, publishers and interest groups attempt to actively engineer content in order to influence what large language models (LLMs) say." Search engine optimisation was a twenty-year war over the first page of results. Its successor is a war over the sentence the model produces when someone asks who to vote for, and it is being fought now, largely unobserved.

The rules exist; the dates keep moving#

Europe has legislated more seriously here than anywhere else, and the legislation is worth reading before praising or damning it. Point 8(b) of Annex III of the EU AI Act classifies as high-risk those "AI systems intended to be used for influencing the outcome of elections or referendum." The Commission's guidance is specific: it covers AI-enabled political ad targeting, chatbots built to persuade voters towards a candidate, and voter advice applications whose personalised recommendations steer preferences towards a party. Campaign logistics, donor prediction and ballot counting fall outside, as may a system that merely checks the tone of a draft leaflet, exempt under Article 6(3)(b) as improving a completed human activity.

That is a carefully drawn line, and it was supposed to bind from 2 August 2026. It does not. Under the Digital Omnibus agreement, reached politically on 6 May 2026 and endorsed by member state representatives a week later, obligations for standalone Annex III systems slipped to 2 December 2027, a delay of sixteen months. What did take effect on 2 August 2026 is the bulk of the Act's other machinery, including the Article 50 transparency duties; labelling obligations for synthetic content carry a transitional deadline of 2 December 2026. The provision aimed squarely at electoral AI is the one part not yet in force, in a year containing the American midterms.

The transparency regulation on political advertising arrived on schedule and produced an outcome nobody in Brussels seems to have wanted. Regulation (EU) 2024/900 — the TTPA — applied from 10 October 2025. Meta stopped selling political, electoral and social issue advertising across the EU four days before, saying the regulation "introduces significant, additional obligations to our processes and systems that create an untenable level of complexity and legal uncertainty." Google followed, complaining that the TTPA "defines political advertising so broadly that it could cover ads related to an extremely wide range of issues that would be difficult to reliably identify at scale." Both had made comparable exits in France, Canada and Brazil.

So Europe's most direct attempt to make political persuasion legible ended with the two largest advertising platforms on earth declining to carry European political advertising at all, and with researchers losing the ad archives that were their best window into paid messaging. Money that cannot buy a labelled advertisement does not evaporate. It moves to influencers, to organic content, to the unlabelled and unarchived.

Four elections, and the awkward results#

Hungary is the strongest case for alarm and, on the numbers, the strongest case against it. Fidesz ran what was probably the most AI-saturated campaign yet conducted in a European democracy; the European Parliament had already warned of the country's deepening rule-of-law crisis in a resolution on 20 November 2025. Survey work summarised by the European Digital Media Observatory found that 73 per cent of Hungarians reported meeting AI-generated political content on social media, that 52 per cent of those who saw a deepfake could not tell whether it was authentic, and that 37 per cent had been misled at least once. Ninety per cent thought such tools unacceptable in politics.

On 12 April 2026, Fidesz lost. Péter Magyar's Tisza party took 141 seats on 53.18 per cent of the party-list vote; Fidesz–KDNP took 52 on 38.61 per cent. Viktor Orbán conceded. Any honest account of political AI has to sit with that result rather than hurry past it.

The Netherlands offers a cleaner measurement of scale, if not of effect. Researchers at the University of Amsterdam, coordinated by the political communication scholar Fabio Votta, monitored campaigns across thirty municipalities from 1 March 2026. Within ten days they had logged 192 posts containing AI-generated images, roughly nine in ten carrying no disclosure of any kind. Lelystad alone produced 46, of which 37 came from PVV candidates.

Canada supplies the most rigorous prevalence estimate available. Victor Livernoche and colleagues at Mila and McGill, in work presented at the ACM Web Conference 2026, analysed 187,778 posts from X, Bluesky and Reddit across 24–29 April 2025, spanning the federal election on 28 April. Synthetic imagery accounted for 5.86 per cent of images overall and 7.9 per cent on X, with right-leaning accounts posting it at nearly twice the rate of left-leaning ones. Almost half the detected material was benign or non-political, and politically harmful synthetic content drew 0.12 per cent of all views on X. Present, in other words, and hardly seen.

The United States is where the practice has become least apologetic. Axios reported on 9 September 2026 that the Texas Senate nominee Ken Paxton had run an advertisement featuring an AI-generated version of his opponent, James Talarico, delivering remarks partly drawn from things Talarico had said and partly stripped of their context. New York officials opened an inquiry into a Republican gubernatorial campaign's AI advertisement said to lack a required disclosure; the campaign called it satire. There is no federal law of substance, and most states require no disclaimer, so the rule governing a national advertisement depends on where the viewer sits. Nate Persily of Stanford Law put the worry at the right level: "The more artificial media that we consume, the less likely that people will be able to discern what is true from what is false."

The liar's dividend, and the evidence against it#

That is the effect worth naming properly. Bobby Chesney and Danielle Citron identified it in volume 107 of the California Law Review in 2019, and their formulation has aged better than most predictions of that vintage. "Liars aiming to dodge responsibility for their real words and actions will become more credible as the public becomes more educated about the threats posed by deep fakes," they wrote. A sceptical public "will be primed to doubt the authenticity of real audio and video evidence." Hence the liar's dividend, which "flows, perversely, in proportion to success in educating the public about the dangers of deep fakes."

The danger, on this account, is that the category of evidence loses its standing. A recording stops settling a dispute and becomes one more contested claim, which is the condition under which a politician caught on tape can decline to be caught. Much contemporary disinformation is not trying to install a belief at all. It is trying to make believing expensive.

There is experimental support for this, with a limit that ought to be reported alongside it. Kaylyn Jackson Schiff, Daniel Schiff and Natália Bueno ran five pre-registered experiments with more than 15,000 American adults, published in the American Political Science Review in February 2025. Politicians who answered a scandal by claiming misinformation gained more support than those who apologised or stayed silent, by 0.17 to 0.21 standard deviations, and the gains reached co-partisans, independents and some out-partisans alike. But that held for scandals reported in text. Against video, claiming a deepfake was largely ineffective. So far the dividend pays best where there is no footage to deny.

The case that the panic is overdone#

A serious body of research holds that the whole frame is wrong, and it deserves more than ritual balance. Start with exposure. Ceren Budak, Brendan Nyhan, David Rothschild, Emily Thorson and Duncan Watts argued in Nature on 6 June 2024 that three assumptions underpinning most public discussion are unsupported: that average exposure to problematic content is high, that algorithms are chiefly responsible, and that social media is a primary cause of polarisation. Fake news websites, they calculate, make up about 0.1 per cent of the American media diet once television is counted. Heavy consumers of untrustworthy material "are often already highly attentive to this content and seek it out across mediums," which inverts the causal story: "exposure levels may reflect people's preferences more than they shape them." If demand is the binding constraint, cheapening supply changes less than it seems.

Then persuasion. Kobi Hackenburg and Helen Margetts, in PNAS in 2024, fed demographic and political data into GPT-4 to generate individually targeted political messages. The messages persuaded. The targeting added nothing: "the persuasive impact of microtargeted messages was not statistically different from that of non-microtargeted messages." The power of current models, they infer, "may reside not in their ability to tailor messages to individuals, but rather in the persuasiveness of their generic, non-targeted messages" — closer to good copywriting than to mind-reading. Against this, Almog Simchon, Matthew Edwards and Stephan Lewandowsky found in PNAS Nexus that personality-matched advertisements did outperform mismatched ones, at a coefficient near 0.03, and estimate 2,490 additional people persuaded per 100,000 exposed. Which finding matters more depends on how many exposures you can buy.

And the sharpest objection: the remedy may be doing the damage. John Ternovski, Joshua Kalla and P. M. Aronow reported in the Journal of Online Trust and Safety in 2022 that informing people about deepfakes "did not enhance participants' ability to successfully spot manipulated videos but consistently induced them to believe the videos they watched were fake, even when they were real." If that generalises, a decade of awareness campaigns has been manufacturing precisely the reflex the liar's dividend requires.

None of this shows the alarm is groundless. It shows that the mechanism is not the one usually described, and that a piece carrying only the alarm is doing advocacy.

Where the philosophy earns its place#

Hannah Arendt's claim about totalitarian propaganda was never that it made people believe falsehoods. Her point was that a sustained assault on factual truth destroys the common world in which people can disagree at all, leaving a population that believes nothing and can therefore be told anything. The Fidesz video attacked no particular fact. It attacked the frame in which facts could be contested.

Reasoning together about common matters, which is what Habermas took a public sphere to be for, presupposes that the reasoners are looking at the same thing. Micro-segmented campaigning is less an abuse of that arrangement than a solvent for it. When no two citizens receive the same message, their disagreement cannot be adjudicated, because they are not disagreeing about anything in common. That is a different failure from polarisation, which at least requires a shared object to be polarised about.

Shoshana Zuboff's argument about behavioural surplus is usually filed under commerce and belongs in politics. Data extracted as a by-product of ordinary use, refined into predictions of what a person will do next, is an asset whose value is indifferent to whether the buyer sells trainers or candidates. The infrastructure was built for advertising and inherited by campaigns, with no adjustment for the fact that a citizen is not a customer.

The Dutch chatbot finding, though, is best explained by Foucault. Power, in his account, does not principally forbid; it produces the visible, the normal and the sayable. A model that mentions local parties in under one per cent of its recommendations persuades nobody of anything. It constitutes a political field in which certain options are not among the options — a machine of visibility before it is a machine of persuasion, and visibility is the more fundamental power.

That leaves Chantal Mouffe, whose distinction between the adversary, whose right to make a case one accepts, and the enemy, who is to be defeated, is the hinge on which democratic conflict turns. A behavioural profile is neither: parameters to be optimised against, not argued with, not accepted, not even hated, merely modelled. The campaign that works this way has not become more hostile to voters; it has left the relation in which hostility or respect could arise. Automating political work has a way of removing the addressee.

The oversight problem, and what is being proposed#

Full Fact's 2026 report, A system under strain, with a foreword by its chief executive Chris Morris, is less apocalyptic than campaigners might wish and more useful for it. Its central observation: "AI-powered search tools, chatbots and automated summaries are becoming routine entry points for political information." Its central worry is the pace of deployment relative to evaluation. It notes concerns about how "increasingly capable, general-purpose AI systems behave under real-world conditions, particularly where oversight and evaluation lag deployment," and that such systems "operate with limited transparency, accountability or regulatory oversight."

Its remedies are institutional rather than technological: strengthen electoral law, impose systemic risk management duties, set standards for AI labelling, open platform data to researchers, and establish "a statutory body to coordinate preparedness and response to threats to the UK's information environment." One may think this too much state or too little. It is at least the right kind of proposal, concerned with who is answerable rather than which model is banned — closer to law written as executable rule than to prohibition.

The academic field is younger than its confidence suggests. Joachim Wiewiura, Amanda Jørgensen and Thomas Ploug published a scoping review in Philosophy & Technology on 21 August 2026, the first, they say, "to investigate the conceptual clarity of notions of democracy invoked in debates on AI." Their finding is uncomfortable for everyone: "Conceptual unclarity and pessimistic views on AI's contribution to democracy often coincide." The gloomiest contributions are the vaguest about what democracy is meant to be.

So the question is not whether AI is good or bad for democracy, which is unanswerable as posed, nor whether the effects are large, which the evidence suggests they are not, at least not yet and not where they were expected. It is which institutions will be answerable for a system that shapes what a citizen sees, and whether anyone will be able to establish afterwards that it did. Hungary suggests electorates are less biddable than the technology's salesmen and its critics both assume. Amsterdam suggests the labelling regime is ignored at nine-tenths scale while the law that would enforce it sits fifteen months out.

Beneath both sits an older worry. A polity that measures its citizens with sufficient resolution acquires, whether or not anyone intends it, the capacity to manage them; and a citizen who is continuously modelled has already had something taken that no disclosure label gives back. Whether that capacity is exercised is a matter of institutions, not of code. On present evidence the institutions are behind, and the ones that are ahead have chiefly succeeded in driving the activity somewhere they cannot see it.