The voice of its clone: AI, fake news and the ballot box on the road to 29N
For the first time in a general election, there is a guest who appears on no ballot paper: artificial intelligence. Can AI change an election? We look at the science, electoral deepfakes, the law… and put Fran to the test with cloned politicians’ voices live on air.
On 5 October Pedro Sánchez called a general election for 29 November, and in his very own institutional statement he already smuggled in the guest who does not appear on any ballot paper: artificial intelligence, “very disruptive”. Videos that were never recorded, audio clips nobody ever said and chatbots that recommend who to vote for. Fran and Carlos ask whether AI can change an election.
What the science says. Two December studies in Nature and Science: more than 100,000 voters in the US, Canada and Poland chatted for a few minutes with a chatbot and support for the candidate rose by two or three points (traditional election advertising moves less than one). What convinces most is not psychological trickery but dropping in facts in full armchair-expert mode… and the more facts the model spits out, the more it gets wrong: the most persuasive AI is the least reliable. Personalised messages do not persuade more than generic ones, and in the Netherlands chatbots recommended almost always the same two parties.
Deepfakes around the world. The fake audio of Slovak liberal Šimečka buying votes, on the eve of polling day; the call with Biden’s cloned voice in New Hampshire, which cost one dollar and ended in a 6 million fine; Suharto brought back to life in Indonesia and 650 million views of AI videos in Hungary. The Alan Turing Institute found no evidence that AI changed any result in 2024, but the real damage is the erosion of trust. And in Spain, across the political spectrum: the fake Sánchez from Junts, the video withdrawn by the PP, García-Page’s cheapfake in Talavera, the 23 YouTube channels with fake news identified by Maldita and 74% of Spaniards worried about not being able to tell true from false.
Who said that? Cloning a voice costs about 6 dollars a month, and in a Queen Mary University of London study we performed no better than heads or tails. Carlos cloned, on his home computer, using the free program F5-TTS and 15 seconds of audio, the voices of Page, Sánchez, Ayuso, Feijóo… and Fran’s own, and then tested it: three clips, one genuine and two fake. Result: three out of four, and the cloned Ayuso scored a goal past him.
The law and the liar’s dividend. Making a deepfake is not forbidden: the European regulation has required it to be labelled since 3 August, the Spanish AI law is still in Parliament and neither will arrive in time for 29N; Brazil bans synthetic candidate content 72 hours before voting. The big danger is that, if anything can be fake, the politician caught out only has to say “that’s AI”, as some defences hinted in the Koldo case. And asking the chatbot who to vote for? According to the UOC, ChatGPT recommended a party in 43% of cases, and Grok is the most tilted to the right. Three tips to finish: be wary of urgency, look for the source and the fact-checkers, and if in doubt, don’t forward it. Hopefully on 29 November it will be the people’s voice that decides, not the clones’.