Marc López Plana described last October 13 in these pages how the networks tied to the Kremlin have entered the Spanish campaign with almost nothing to invent. They seize the housing crisis or what happened this summer in Ceuta and stitch them into a narrative of democracy in a governance crisis. He calls it narrative absorption, and rightly warns that identifying amplification networks does not prove they shift a single vote. The discontent over housing exists without Moscow’s involvement.
A detail from his analysis opens the clearest doorway to understanding what lies ahead: Pravda ES redistributed material from the portal La Bandera based on 40dB surveys. Nobody had to falsify the poll; in this case it was enough to lend it its channel. I have spent years explaining to campaign teams that a survey can excite or discourage, but the worst-case scenario is when it is ignored. Today there is a worse one to add: the survey that no one conducted and that everyone shares.
Election misinformation has for years been discussed as a hoax problem. What is being manufactured now are the signals we use to decide what to believe. Maldita.es documented on October 16 a batch of TikTok graphics with supposed polls for the 29N, all predicting a PSOE victory. There were September CIS figures presented as current, along with a seat distribution attributed to the CIS (which does not publish seats in its barometers) and graphics labeled as AI-generated, one of them with percentages changing in seconds.
“Election misinformation has for years been discussed as a hoax problem. What is being manufactured now are the signals we use to decide what to believe”
Two weeks earlier there had circulated an manipulated image, in the Data Hive format and attributed to La Razón, which gave the PSOE an absolute majority; the company clarified that someone had altered one of its images. Nor is this entirely new. On the day of reflection ahead of the April 2019 general election, a poll attributed to the Diari d’Andorra circulated that newspaper had not published, as documented by a study by Paniagua, Seoane and Magallón-Rosa of the think tank CIDOB. Moreover, it was credible because during the vote-curbing period polls are published across the border.
In Brazil, the agency Lupa published on October 15 the comparison of its verifications across the first rounds of presidential elections in 2018, 2022 and 2026. Claims challenging the integrity of the electoral system have fallen from 23.4% in 2022 to 7.5% this year. Instead, there is growth in content that fabricates signals of political strength, such as endorsements, rejections, or crowds that never existed.
They appear in seventeen of the eighty verified network contents in 2026, or 21.3%, up from 14.9% in 2022 and 15.4% in 2018. In fifteen of those eighty there was material produced or altered with artificial intelligence, something Lupa had not found in the two prior cycles. Targeting the ballots collides with courts and fact-checkers. Creating the impression that a candidate is sweeping the board is far harder to debunk, because it acts on something no one certifies: the atmosphere.
In the strategy rooms in which I have worked, on both sides of the Atlantic, the purpose of a false poll is well understood. The convinced voter ignores it. Its target is the undecided voter who wants to back whoever is perceived to be winning, the donor who doubts, the journalist choosing who to interview, and the rival’s cadres who begin to distrust their own candidate, and that is why it works even when it is denied: when the denial arrives, many have already made up their minds.
“Fabricating the feeling that a candidate is sweeping the board is much harder to debunk, because it acts on something that no one certifies: the mood”
If in Spain pro-Russian networks amplify what others publish, in Latin America a Russian operation has already taken the next step and manufactured the source. OpenAI reported on October 8 that it had dismantled Dark Clark, which sought to damage Ukraine’s reputation in the region and meddled in Argentina and Bolivia’s domestic politics. It operated through a supposed research platform, the Social Research Center, and published more than sixty articles with the help of regional people who were acting in good faith without knowing for whom.
It is the first operation that OpenAI assigns to category 5 of its six-point scale, although the company warns that the operators exaggerated their effectiveness. In the same report it describes an Iranian operation that placed about a hundred articles, signed by non-existent journalists, in a dozen digital outlets. The architecture repeats itself and costs little: a neutrally named research center, signatories that lend it the appearance of research, and real media outlets that end up sheltering it.
That material ends up in AI assistants, increasingly the first stop for many citizens instead of a newspaper. On September 20, the day of regional elections in Berlin and Mecklenburg-Vorpommern, consultant Pavel Gurov repeated ten politically charged questions to the major assistants. Perplexity cited Russia Today in twenty of every hundred responses; ChatGPT and Gemini, in none. The sample is small, though reproducible, and eight days later Perplexity announced it had blocked those domains.
“The architecture repeats itself and is cheap: a neutrally named research center, signatories that lend it the appearance of research, and real media outlets that end up sheltering it”
The mechanism behind all of this is simple. If the question arrives framed as part of a deceitful narrative and there are no reliable sources to answer it, the assistant gathers the pages that share that framing, and the one who produces volume with the appearance of a source gains prominence in the reply. In Spain, a study by the UOC published in March found that ChatGPT gave a voting recommendation in 43.2% of inquiries asking for a vote. It’s another problem, but it shows that the assistant is already part of the electoral conversation.
How does Spain defend itself? It has had a procedure for countering disinformation since 2020 and a coordination network within the Interior Department that is activated for every election since 2019. Brussels has issued guidelines to the major platforms and its political advertising regulation has been in force since October 2025. Almost everything is designed for identifiable hoaxes or for paid ads. A false poll circulating on TikTok or a front center cited by an assistant are left in limbo, and no party has publicly, at least to our knowledge, a protocol for this threat.
The teams remain organized to debunk piece by piece, something necessary that falls short when what is manufactured is a climate. It is also essential to monitor the supporting signals that circulate, document each incident, and scrutinize sources that suddenly appear cited, especially if they are recent and of diffuse authorship. Knowing what the assistants respond about a candidacy is becoming as elementary as reading today’s press, and it will apply equally to the municipal and regional elections in May.
The most uncomfortable aspect concerns those of us who practice this. The same tools are within reach of any campaign team, and there will be those who think that a timely invented poll is worth more than ten real ones. It is a moral error and also a miscalculation. Credibility is built slowly and destroyed by a single disclosure, and fact-checkers and researchers track these patterns with increasingly fine methods. Our craft should say so aloud before others have to tell us.
“Credibility is built slowly and destroyed by a single disclosure, and fact-checkers and researchers track these patterns with increasingly fine methods”
On Tuesday, November 24, the blackout period begins. For five days it is prohibited by law to publish polls in Spain, and polls published in Andorra will circulate alongside those masquerading as them. Anyone who has not prepared their team in advance to distinguish them will be late, because that week the denial always lags behind.
