When AI whispers to us what to vote

On November 29th we will vote in a general election in which part of the electorate will have prepared their vote by conversing with an artificial intelligence. Disinformation no longer reaches us only through viral deepfakes, but through a more intimate whisper: the private, synthetic, and witness-free conversation we have with chatbots.

The data is still modest, but it is growing. According to the 2026 Digital News Report, 8% of Spaniards already use chatbots to stay informed every week, although only 18% trust them. In the United Kingdom, one in eight voters turned to them in 2024 to guide their vote. And there is a key figure: 1.6 million people will vote for the first time in a general election. It is a generation that uses AI for almost everything and arrives at the polls with a volatile vote that parties are fighting over with sudden bursts of passion.

The issue is that these models are not designed to make us better citizens, but rather to be efficient, to tell us what we want to hear, and to keep us from wanting to log off. And a confidant who always agrees with us does not help us think; it traps us even further in our own prejudices. What could go wrong if democracy is precisely about knowing how to converse with those who do not think like us? The power of AI in electoral processes is not prescriptive and occasional, but rather interpretive and constant: it does not tell us who to vote for, but it frames how we understand what we vote for. With November 29th approaching and the municipal elections in the spring, the question is no longer whether we will use AI to inform ourselves, but what it implies that AI also conducts covert campaigning.

A few days ago, Anthropic, the company behind Claude, published how it had prepared for the United States midterm elections, including official information on how and where to vote, new systems to detect manipulation campaigns, and political bias tests reviewed by external organizations. To begin with, the figures provided were surprising: 5.75% of conversations with Claude in the US already touch on politics, but only 0.01% directly ask who to vote for. Conversely, according to a survey by the AAPC Foundation, one in five American voters has already asked AI about elections and politics, and nearly four out of ten receive AI-generated electoral responses in search results, half of them without knowing it. In the Netherlands, the Data Protection Authority has found that ChatGPT, Gemini, Mistral, and Grok end up recommending parties even when not asked to.

In Hungary, before this year's elections, the NGO Liberties gave ChatGPT and Gemini voter profiles that matched exactly the platform of each party. Profiles aligned with Fidesz, Orbán's party, were recognized without issues, while those aligned with Tisza, the opposition party, were often misclassified or ignored.

The safeguards announced by Anthropic have been designed for the United States, Israel, and Brazil. However, companies often apply the same policies everywhere, and what a chatbot tells us about politics can reflect pressures from governments here and there. This is explained by Jordi Calvet-Bademunt, a researcher at The Future of Free Speech, the organization that, precisely, has tested Claude on behalf of Anthropic with hundreds of questions on controversial topics. From India to Turkey, many governments use national security laws to influence what chatbots say about those in power. The EU is also intervening: in August, it placed ChatGPT at the most demanding level of the Digital Services Act (DSA), with the obligation to mitigate electoral risks.

However, bias tests can compare left and right or liberal and more conservative orientations, but they leave out the national axis. And the problem is deeper than the axes: Accent Obert warns that chatbots speak Catalan, but they operate with cultural frameworks learned far from here, mainly Anglo-Saxon. Using our language is not equivalent to understanding the framework or the Catalan-speaking complexity.

The risk, then, is that we delegate our judgment. And this is a double-edged problem: on one hand, we gamble with our critical thinking; on the other, we may end up unintentionally assimilating the frameworks that AI silently imports and which have little to do with our political reality. Researchers Wencheng Lu and Zhenni Hu call it cognitive disqualification: when we delegate judgment, we lose the muscle to exercise it. Their proposal is a Socratic chatbot, which does not answer but asks questions.

And it is not science fiction: there are already some experiences. The European project TITAN has tested with citizens in Belgium, Denmark, Italy, and Bulgaria an assistant that, when faced with a suspicious news item, does not tell you if it is false, but helps you discover it. Almost two out of three Belgian students said it made them think more critically. In Germany, a group of students created wahl.chat, a chatbot that answered questions based on official programs.

Can you imagine if we could have a public one, in Catalan and situated, that instead of telling us who to vote for or how to verify news, helped us investigate why we think what we think? And what if, in addition, we designed it with young people? The mere process of creating it would be a dual exercise: of democratic literacy and technological sovereignty.

Because the question is not whether AI enters the campaign, but with what logic it does so. A neutral AI does not exist, but we can demand one that is fair, responsible, and transparent. And we do not need an AI without friction either, but one with the right friction: that which makes us question ourselves and pause to decide better. And then the democratic triumph of the next elections will be that, when we arrive at the polls, the decision will be reflected upon and radically our own.