Young people gathered outdoors embody a generation already familiar with AI assistants. The image accompanies an investigation into political information, bias, and electoral trust. Credit: Eliott Reyna / Unsplash.
The 2027 French presidential election is shaping up to be the first national vote in the age of generative AI assistants. These tools could serve as political intermediaries at scale. Voters are already consulting them to understand platforms. Researchers and regulators, meanwhile, are warning about bias, hallucinations, political deepfakes, and election profiling.
A New Political Intermediary
Artificial intelligence does not vote. But it can sort information, rephrase a platform, and summarize a public figure. It also gives its answers the appearance of technical neutrality. That shift is what a report published by Franceinfo on July 3, 2026, illustrates. The piece focuses on the possible uses of AI in the 2027 French presidential campaign.
The report describes young voters who are already asking ChatGPT or other assistants to help them navigate the political landscape. It also recounts the case of a prototype using a voice resembling Gabriel Attal’s, with answers that were sometimes inaccurate or made up. According to Franceinfo, the former prime minister’s team was not informed of this experiment and did not wish to comment.
This point is crucial: the issue is not just the spread of a spectacular fake. AI’s influence on voting can also come through more ordinary details. It lies in the sources selected, the order of the arguments, or the way candidates are named. It can also come from a nuance left out or the confidence created by a smooth answer.
Already Measured Uses
The available figures do not prove that AI can swing an election. They do show, however, that the political use of these tools is no longer marginal. An Ipsos BVA survey for the Jean-Jaurès Foundation was published on February 5, 2026. It found that 48% of French people have already used or are considering using generative AI to learn about politics.
In that same survey, 25% of respondents say they have already asked an AI model about a political figure or party. Another 23% have not done so yet but say they are willing to try. The figure rises to 75% among 18- to 24-year-olds. For the 2027 presidential election, 28% of French people say they plan to use it this way.
The March 2026 municipal elections provide a first point of observation. A memo published by Terra Nova uses Toluna Harris Interactive data for M6 and RTL. According to the memo, 11% of the voters surveyed say they used AI to get information about their town’s municipal campaign. The survey was conducted online on March 15, 2026. It covers 4,145 people registered on the electoral rolls in municipalities with 3,500 residents or more.
Public Sénat repeated and expanded on these results. The network notes that 16% of those surveyed used AI to help guide their vote. Among them, 7% wanted to confirm a choice, 5% had changed their minds, and 4% were trying to break a deadlock. Among those under 25, that figure reaches 35%. These numbers are still a minority, but they shift the debate. Generative AI in elections is no longer just a campaign tool for political teams. It is also becoming a consultation tool for voters.
Bias Does Not Need To Give An Order
Talking about biased AI does not mean a piece of software has a simple, stable, and explicit electoral preference. Algorithmic bias can stem from incomplete training data, internal instructions, or unequal access to sources. It can also come from tone choices or an opaque ranking mechanism. An answer may seem balanced while favoring certain topics, certain wording, or certain players.
For a non-specialist reader, the difficulty lies in the chatbot’s very form. An AI response is often clear, personalized, and instant. It can cite accurate elements, mix in others, and then conclude with confidence. Users would not necessarily grant that same trust to a flyer or a partisan post. The error does not have to be massive to matter. A poorly summarized platform, a measure attributed to the wrong candidate, or an incomplete comparison can change the perception of an uninformed voter.
Caution is therefore essential. The Ipsos, Terra Nova, and Public Sénat data describe uses, intentions, and statements. They do not measure a causal shift in the election caused by AI. Still, they are enough to raise a democratic question. If a growing share of citizens turn to conversational assistants to understand politics, who controls the quality of those answers? Who ensures their traceability and balance?
Deepfakes, Voice Cloning, and Profiling
The most visible risk remains election deepfakes. They can take the form of an image, a video, or audio that imitates a person. The goal is to make them say or do something they did not say or do. On this point, the CNIL has already issued warnings. In a Localtis article published on June 20, 2024, the authority highlighted several risks. It cited hallucinations, biased content, hyperfakes, labeling AI-generated content, and voter profiling.
The CNIL fact sheet on hyperfakes, updated in February 2026, defines these contents as images, sounds, or videos. They are created or altered by AI. The fact sheet also notes that a person’s voice or image can be misused without their consent. The consequences can affect their privacy, reputation, or identity.
In a presidential campaign, these risks are not limited to the final week. A fake audio clip can be quickly debunked, yet still leave a mark. A candidate’s assistant can respond to thousands of voters. Its corpus, priorities, or limitations are not always visible. A party can also sharpen its targeting using personal data. This point brings the debate on presidential artificial intelligence closer to the one on data protection.
What Safeguards Before 2027?
The first safeguard is editorial: clearly attribute facts and distinguish usage from influence. Saying that voters use AI does not mean AI will decide the election. Saying that a prototype exists does not mean it should be presented as an official campaign tool. That line matters because it prevents a real trend from being turned into technological panic.
The second safeguard is political and regulatory. Campaigns will need to specify when they use AI-generated content. They will also need to explain how they label artificial creations and what data they collect on voters. Platforms, for their part, will need to make sources and limitations more visible when their responses touch on civic information.
The third safeguard concerns citizens. Faced with a political ChatGPT, Gemini, Mistral, Claude, or Grok, the right question is not just “What does the tool say?” It is also more practical: where is the information coming from? What is missing? What primary source can be checked? Does the answer distinguish facts from interpretations?
The 2027 presidential election will probably not be decided by a machine. But it could be told, filtered, and sometimes distorted by assistants consulted privately, at scale. That is where the quiet risk of biased AI lies. Not in a visible voting instruction, but in the day-to-day mediation of political information.