AI in Elections and the Challenge to Information Integrity

AI in Elections and the Challenge to Information Integrity is the civil-society chatbot-reliability research report published by AfricTivistes on 22 April 2025 in partnership with Berlin-based Democracy Reporting International (DRI), with the English-language PDF release following in June 2025. The report tests five major generative-AI chatbots against questions on the 17 November 2024 Senegalese legislative elections and supplies the corpus's first sub-Saharan-African civil-society empirical evidence base on whether chatbots are fit to serve as electoral-information sources. The original French announcement carries the title Intelligence artificielle et Élections : fiabilité et intégrité de l'information en question !.

Method

The study was conducted in October 2024 — the month preceding the 17 November 2024 Senegalese legislative elections — and tested five generative-AI chatbots: Microsoft's Copilot, Google's Gemini, OpenAI's ChatGPT-4o and ChatGPT-4omini, and Anthropic's Claude. The chatbots were asked the same set of 14 questions in French and Wolof about the Senegalese legislative elections, and their responses were evaluated against three criteria: accuracy (correspondence with official sources), consistency (uniformity and absence of contradictions across repeated queries), and completeness (the depth and relevance of the responses produced).

The bilingual French–Wolof design is one of the report's methodologically distinctive contributions: it brings into the chatbot-audit literature an explicit test of low-resource-language electoral-information delivery, where Wolof — the principal vernacular language of Senegal and a working language of Senegalese public discourse alongside French — sits outside the high-resource-language training-data concentrations that the dominant chatbot families are optimised for. The methodology transposes Democracy Reporting International's standing chatbot-audit programme — first applied to the 2024 European Parliament elections in ten European languages, and subsequently extended to the German federal elections and the follow-up EU chatbot-companies study — onto the Francophone African electoral context, with AfricTivistes leading the partnership and DRI collaborating as co-researcher.

Findings

The report's headline finding is that chatbots cannot be considered entirely reliable sources of electoral information. The per-tool performance breakdown is differentiated: ChatGPT-4o and Claude showed relatively accurate responses but were limited to general information; Copilot and Gemini demonstrated significant shortcomings with frequently outdated or incomplete responses; ChatGPT-4omini exhibited the weakest performance with the highest error rate. Cutting across the per-tool results, none of the chatbots could integrate real-time updates of the Senegalese political context — a structural rather than per-vendor limitation that the report identifies as the binding constraint on chatbots' fitness for electoral-information delivery.

The report's principal critical move follows from a finding that is sharper than the headline rate of incorrect responses: some chatbots provided incorrect information with high levels of confidence, posing a serious risk of disinformation, especially during electoral periods. The combination of confident framing and substantive inaccuracy makes the chatbots' confidently-wrong outputs a higher-impact disinformation surface than transparent-uncertainty outputs would be — a citizen asking a chatbot about a candidate, a polling location, or an electoral procedure and receiving a confidently-worded incorrect answer is more likely to act on it than a citizen who receives a hedged or uncertain response. In the report's analytical frame this confident-but-wrong output pattern is the proper object of the electoral-information-integrity concern rather than chatbot inaccuracy in the abstract.

Recommendations and posture

The report's recommendations are addressed to citizens, journalists, and civil-society actors and concern the practical posture that publics should adopt towards chatbot output during electoral periods. The canonical recommendation is that citizens, journalists, and civil-society actors should systematically verify information obtained via these tools by cross-referencing them with official and recognised journalistic sources, and that public-awareness work on the limitations and best practices of chatbot use is essential to guarantee responsible information consumption. The longer-term recommendations extend to the structural register: heightened caution in generative-AI use, strengthened digital literacy, algorithmic transparency, and the accountability of tech platforms for the electoral-information outputs their tools produce in low-resource-language and Global-South contexts.

The report is positioned inside AfricTivistes's Media and Information Resilience pillar of the 2025–2029 Strategic Plan and operates as the analytical anchor for the wider AI-and-elections programme line. The operational counterpart is the €175,000 Election Civic Tech Fund supporting civic-technology initiatives across fourteen African countries, where the chatbot-reliability finding supplies the evidence base for civic-tech investment in alternative electoral-information channels not dependent on chatbot reliability. The report's findings also feed downstream into AfricTivistes's Foreign Information Manipulation and Interference research line, where the chatbot output pattern is treated as one vector inside the wider electoral-information-manipulation landscape that the FIMI evidence base is building out.

Posture within the corpus

Within the corpus, AI in Elections and the Challenge to Information Integrity is the first publication-class entry anchored on AfricTivistes, the first publication anchored on Francophone West Africa, and the first publication in any geography on chatbot reliability as an electoral-information-integrity question. The publications slate has previously carried the Anglophone-rooted African anchor in the Paradigm Initiative-led Legal Analysis of Laws, Policies and Government Strategies Relating to Artificial Intelligence in Kenya, Mauritius, Rwanda, South Sudan, Tanzania, Uganda and Zambia, the EngageMedia-led Governance of Artificial Intelligence (AI) in Southeast Asia, the Derechos Digitales-led Decisiones automatizadas en la función pública en América Latina, and the APC / SMEX regional reports on Global South AI governance; the AfricTivistes-DRI study installs the Francophone-West-Africa anchor that the corpus's African coverage previously lacked and gives the publications slate its first non-Anglophone Sub-Saharan African anchor.

The publication is also the corpus's first instance of the chatbot-audit methodological register: where the existing Global South publications slate has clustered around legal-analytical surveys (the Paradigm Initiative seven-jurisdiction East-and-Southern-African legal analysis), sectoral audits (Velasco and Venturini's five-dimension human-rights audit in the Derechos Digitales case), and civil-society-participation analytical frames (Jun-E Tan's ASEAN-wide analysis in the EngageMedia case), the AfricTivistes-DRI study supplies the first empirical-audit register — a small-N, criteria-driven, repeated-query evaluation of named commercial AI products against named electoral facts. The methodological lineage runs through DRI's chatbot-audit programme on the 2024 European Parliament elections and the German federal elections; the Senegal study is the first transposition of this methodology to a sub-Saharan African electoral context and the first instance to add a non-Latin-script low-resource language (Wolof) to the methodology's language stack. The structural shape of Global-South-civil-society anchor publishing in partnership with a Global-North methodology-bearing organisation is also new on the publications slate: prior Global South anchors have been single-publisher or all-Global-South-author productions, where this report is a North–South methodological collaboration with the Global South organisation in the publisher and operational-lead seat and the Global North organisation contributing methodological scaffolding.

The publication is referenced from the AfricTivistes body's AI, elections, and information integrity section, where the chatbot-reliability study is the leading-edge empirical contribution that anchors the wider AI-and-elections programme line through the 2024–2025 cycle and into the 2025–2029 Strategic Plan.