AI chatbots deliver inaccurate election advice, raising concerns about electoral interference

Election voting advice from AI chatbots 'inaccurate and unreliable'

Research during Hungary election shows AI recommended parties not running and gave highly volatile answers to identical prompts AI chatbots provide inaccurate, inconsistent and unreliable guidance to voters asking which party they should back, a study suggests, often recommending the wrong party, not mentioning the right one, or listing parties not even running. "The results raise serious concerns about the reliability of general-purpose AI systems in electoral contexts," the study, published by the civil liberties group Liberties and based on research during this year's Hungarian parliamentary elections, concluded. "They misclassified profiles, omitted relevant parties and included parties not running in the election," said the report, shared with the Guardian. Outcomes were also highly volatile, with "materially different" answers given to identical prompts. The report's most striking finding from the Hungarian election - convincingly won by Péter Magyar's opposition Tisza party - was an overwhelming bias in party visibility, particularly as regards Tisza: in 90% of cases when ChatGPT was fed a detailed Tisza-aligned voter profile, it failed to recommend the party. Magyar's landslide brought to an end Viktor Orbán's 16-year grip on power as prime minister and head of his national-conservative Fidesz party. The researchers noted that unlike voting advice sites or independent media explainers, general-purpose AI (GPAI) systems did not disclose how they generated political guidance, did not yield reproducible results, and were not subject to election-related public oversight. However, their responses often look confident and authoritative, making users more likely to trust them. With 29.8% of Hungary's population identified as AI users, the researchers tested the advice given by the two most popular, ChatGPT and Gemini. Based on party positions outlined on Voksmonitor, a respected Hungarian voting advice app, Liberties created five distinct voter profiles, each aligned with one of the five parties registered on national lists for Hungary's parliamentary elections. Each voter profile was then tested 10 times in ChatGPT and 10 times in Gemini, using two different requests: one for direct advice on which party the fictitious voter should back, and another for a percentage match with each of the five parties. In percentage-matching tests, ChatGPT assigned Tisza a score in just 2% of cases. Instead, users with Tisza-aligned views were routinely recommended a smaller party unlikely to cross the 5% threshold for parliament, or parties not on the national ballot. By contrast, Fidesz-aligned voter profiles were recognised far more consistently. In direct advice prompts, ChatGPT identified Fidesz as the single party the user should vote for in about 50% of cases, presenting it as a primary option in the remainder. The analysis does not claim that the outcome of the election was affected: Tisza still won decisively. But, the researchers warn: "In a more competitive election or where voters are less certain, such outputs could ... potentially impact the outcome." Other issues included inconsistency, with the same voter profile matched with radically different parties in successive tests, and the inclusion - in 96% of responses from ChatGPT and Gemini - of parties not on the 2026 ballot. The researchers also noted that both AI models regularly began their responses with polite disclaimers saying they "cannot give political advice" - before providing several paragraphs of dense, highly persuasive party recommendations. "The answers appeared well-argued, precise and authoritative," the researchers said. "This creates a risk that users may treat the outputs as reliable, even though the underlying method is opaque and the results are unstable." The report argues that the underlying cause of the errors probably lies in training data gaps, filters and language-processing limitations. Tisza surged to prominence only after 2024, meaning static AI training models struggled to place it on their map. But they said the findings exposed a serious regulatory gap: the EU's AI Act obliges providers of GPAI models to assess systemic risks, and its Digital Services Act covers "systemic risks" to electoral processes - but AI chatbots fall between the two. Liberties said safeguards should be developed for AI systems providing political advice, and AI providers should stop offering personalised voting recommendations unless they could guarantee transparency, accuracy, consistency and accountability. "General-purpose AIs should not present opaque and unstable political matching as if it were reliable electoral guidance" for voters, said Eva Simon, the head of Liberties's tech and rights programme. "Democracy cannot rely on opaque systems that claim neutrality while delivering advice they cannot explain, reproduce or guarantee to be accurate."

Yet another study says AI is bad for elections, and the rabbit hole gets worse

AI may be an election bomb waiting for someone to light the fuse Another election, another study has concluded that asking an AI chatbot for political guidance is a spectacularly bad idea. Research conducted during Hungary's 2026 parliamentary election found that ChatGPT and Google Gemini offered not just inaccurate but also inconsistent and unreliable voting advice. The chatbots misclassified voter profiles, overlooked relevant parties, recommended parties absent from the ballot, and sometimes produced materially different answers when given the same information repeatedly. AI keeps failing the voter test The Civil Liberties Union for Europe created five fictional voter profiles based on the positions of the five parties contesting Hungary's national lists. Each profile was tested repeatedly through prompts asking for direct voting advice and percentage-based party matches. Recommended Videos ChatGPT failed to recommend the opposition Tisza party in 90% of tests involving a detailed Tisza-aligned profile. During percentage-matching tests, it assigned Tisza a score in only 2% of responses. Fidesz-aligned views were identified considerably more consistently, while 96% of responses from ChatGPT and Gemini included at least one party that was absent from the 2026 ballot. The researchers found no evidence that this imbalance was deliberately engineered, and they did not claim that chatbot responses affected the election result. Training-data gaps, safety filters, language-processing limitations, and Tisza's rapid rise after 2024 could all have contributed. Those explanations offer little comfort to voters receiving polished and persuasive recommendations from systems whose reasoning remains opaque. The warning joins a growing stack of similar research. During Scotland's 2026 election, Demos tested five AI services with 75 election-related questions and found factual errors in 34.1% of responses. ChatGPT produced errors 46.2% of the time, including incorrect election dates and eligibility rules, invented candidates, and fabricated political scandals. Almost half of the responses provided no citation or supporting link. A Dutch regulator reached another worrying conclusion in 2025. Despite the country's broad multiparty system, four tested chatbots directed voters toward only two prominent parties in 56% of interactions. Separate academic investigations have also found recurring political preference patterns across ChatGPT and Gemini, although the direction and severity can change between models, prompts, languages, and elections. Politicians are already learning to bend the answers Accuracy and embedded bias represent only part of the problem. The information feeding these systems can also be deliberately shaped. The New York Times recently reported on Missouri political candidate Dustin Lloyd, whose priorities were barely reflected when voters asked chatbots about him. Lloyd published a carefully structured question-and-answer page on his campaign website. Subsequent chatbot responses began connecting his personal background with his policy goals, showing how quickly a campaign can alter the AI-generated version of its candidate. Maintaining an accurate campaign website is hardly sinister. However, the same mechanism creates an obvious opening for exaggerated claims, attack pages, fake organizations, and sites designed primarily to influence AI answers. A BBC investigation demonstrated how low the technical barrier can be. A journalist spent roughly 20 minutes publishing a fabricated blog post declaring himself the world's finest hot-dog-eating technology reporter. Within 24 hours, ChatGPT, Gemini, and Google's AI Overviews were repeating parts of the invented story. Though Claude resisted the bait. Similarly, another BBC report explored Google's attempts to combat this emerging manipulation industry. The company has documented websites containing instructions designed to hijack browsing AI systems, influence recommendations, promote particular businesses, and potentially steal data. Google expects these indirect prompt-injection attempts to grow in scale and sophistication. Researchers have also shown that AI-enhanced search engines remain susceptible to specially developed manipulation techniques. Strategies including rewritten-query stuffing and splitting promotional text into segments doubled the manipulation rate compared with a baseline attack in one 2026 study. Google has since expanded its spam rules to cover attempts to distort answers generated by AI Overviews and AI Mode. An election bomb with an unpredictable blast radius AI chatbots combine several volatile qualities. Their answers can be inaccurate, politically uneven, and persuasive. These are also difficult to reproduce and dependent on websites that campaigns or hostile actors can modify. The risk extends far beyond a chatbot explicitly endorsing the wrong candidate. Electoral damage could begin with an omitted party, outdated voting instructions, a fabricated scandal, or a strategically planted page that gets laundered into an authoritative-sounding answer. AI has become an election bomb, with politicians, platforms, researchers, and opportunistic manipulators having their hands on the trigger-and the blast radius is unknown.