# If Voters Are Going to Use Artificial Intelligence, Give Them One They Can Trust **Date de l'événement :** 04/09/2026 * Publié le 04/09/2026 ### Date 04/09/2026 ## Chapô **As artificial intelligence becomes an increasingly common source of political information, its use in elections raises questions not only about misinformation and manipulation, but also about how it might help voters make more informed choices. Jonne B. Kamphorst, Assistant Professor of Political Science at Sciences Po, discusses findings from the study “[A nonpartisan source-grounded AI voter guide is perceived as trustworthy and affects voting intentions](https://osf.io/preprints/psyarxiv/whsm8_v1)”, co-authored with Joseph S. Mernyk, Keshav Sivakumar, Adam Bonica and Robb Willer. If voters are already turning to AI for electoral information, how can democracies ensure that the systems they use are transparent, non-partisan and worthy of their trust?** ## Corps du texte Most discussion of artificial intelligence and elections has focused, understandably, on what A.I. might do to democracy. Deepfakes can make candidates appear to say things they never said; generative A.I. can make political misinformation cheaper to produce and harder to identify; and large language models can still invent facts or produce answers whose origins are difficult to reconstruct. Faced with these risks, pundits and technology companies have generally treated elections as an area in which their models should tread carefully. Ahead of the 2024 U.S. election, for instance, both OpenAI and Anthropic placed restrictions on how their systems would handle certain questions about candidates and voting. ### An opportunity to make it substantially easier for voters to understand candidates There is another possibility, however, that has received far less attention. Used differently, the same technology could make it substantially easier for voters to understand what candidates believe, what they have done in office and what they are proposing to do next. The information needed to cast an informed vote is often scattered across party platforms, candidate websites, legislative records, interviews and news coverage, and assembling it can require more time than most voters can reasonably devote to the task. Large language models are unusually well suited to taking a complicated body of information, translating it into ordinary language and allowing people to interrogate it through their own questions.  The difficulty is that elections place unusual demands on an A.I. system. If voters are going to rely on one for political information, it should be possible to know what information the system is drawing on and what rules govern its answers. Those requirements sit uneasily with the incentives of the companies developing today's leading models, which have commercial reasons to protect training data, system designs, and other proprietary information. In many applications, that opacity may be an acceptable price for using a powerful technology. In an election, where even the suspicion of concealed political influence can undermine trust, it is a serious problem. Yet voters are unlikely to wait for this problem to be solved before turning to A.I. for political information. In a survey my colleagues and I conducted in the United States between April and August 2025, 13.9 percent of respondents said they had already asked an A.I. chatbot about candidates or questions on their ballot. They were doing this even though there was no widely used A.I. voter guide designed for that purpose; they were simply putting election questions to general-purpose systems. The relevant policy choice, in other words, is not between a future in which voters use A.I. and one in which they do not. It is between voters relying on systems that were never designed to function as non-partisan audable election guides and giving them an alternative that is. ### One advantage of conversational A.I. over more conventional forms of election information: voters can begin with the issues they care about In the final week of the 2024 election, my colleagues Joseph Mernyk, Keshav Sivakumar, Adam Bonica and Robb Willer and I tested what such an alternative might look like. We built a chatbot on top of GPT-4, but instead of allowing it to draw freely on the internet or on everything contained in the model's training data, we restricted it to a defined set of non-partisan source material: Ballotpedia pages covering federal and state-level candidates in California and Texas. Ballotpedia is a nonprofit political encyclopedia with an explicit neutrality policy and published editorial procedures, and we separately instructed the model to remain politically neutral. We then conducted a preregistered experiment with 2,474 eligible voters who had not yet cast their ballots. Half were invited to use the chatbot for at least two minutes; the other half were asked to consult whatever sources they would ordinarily use when deciding how to vote. Participants used the chatbot primarily to ask about policy issues, which they raised about four times as often as questions about particular candidates. That pattern suggests one advantage of conversational A.I. over more conventional forms of election information: voters can begin with the issues they care about and ask how the candidates on their ballot compare. Just as importantly, participants regarded the information they received as trustworthy, accurate and unbiased, and Republican participants did not perceive the chatbot as favoring Democrats. Among the least politically knowledgeable third of the sample, access to the tool also increased confidence in their ability to make an informed electoral choice, while participants more broadly became more supportive of candidates whose ideological positions were closer to their own.  ### A good AI election-information system should help voters understand which choices on the ballot most closely correspond to the preferences A good AI election-information system should help voters understand which choices on the ballot most closely correspond to the preferences they already hold. But if such systems are going to become part of elections, their informational foundations should neither depend entirely on the private decisions of technology companies nor be controlled by the government of the day. A better model would be an independent, publicly funded institution - more like a central bank than a government department - responsible for maintaining a trusted source of political information and making it accessible to voters. Its mandate should be narrow. Rather than deciding which political arguments are correct or which policies are desirable, it would keep, in one place, a reliable and auditable record of political activity: roll-call votes, speeches, interviews, party manifestos, official statements, legislative decisions and other clearly sourced material. The rules determining what enters that record should be public, and the sources behind every piece of information should be visible. The institution could then build an open-source A.I. system that allows voters to query this material in ordinary language, while making it possible for researchers, journalists and political parties to scrutinize both the underlying information and the way the system uses it. The value of such an institution is likely to grow as deepfakes and A.I.-generated political misinformation become more convincing and harder to distinguish from authentic material. Voters are already beginning to use A.I. to understand elections, and asking a chatbot about an unfamiliar candidate may soon become as ordinary as using a search engine is today. The choice for democracies is therefore not simply whether to allow A.I. into elections or try to keep it out. It is whether voters will have to rely on systems never designed to deserve their political trust, or whether we will build an alternative that is. We spent considerable effort thinking about how to protect voters from artificial intelligence. We should think just as seriously about how artificial intelligence might work for them. ### Thématique `#Numérique` **Licence :** `#CC-BY-ND (Attribution, Pas de modification)` **Langue :** `#Anglais` --- ### Navigation pour IA - [Index de tous les contenus](https://conference.sciencespo.fr/llms.txt) - [Plan du site (Sitemap)](https://conference.sciencespo.fr/sitemap.xml) - [Retour à l'accueil](https://conference.sciencespo.fr/)