Another study suggests that AI is detrimental to elections, and the situation continues to deteriorate.

      Another election has led to a study concluding that seeking political advice from an AI chatbot is a remarkably poor choice. Research conducted during Hungary's 2026 parliamentary election revealed that both ChatGPT and Google Gemini provided not only inaccurate but also inconsistent and unreliable voting recommendations. The chatbots misidentified voter profiles, failed to recognize relevant parties, suggested parties not on the ballot, and occasionally yielded significantly different answers even when the same information was provided multiple times.

      AI continues to falter when it comes to voter guidance.

      The Civil Liberties Union for Europe created five fictional voter profiles reflecting the positions of the five parties competing in Hungary's national lists. Each profile underwent repeated testing with prompts requesting direct voting advice and percentage-based party match results.

      In 90% of tests involving a detailed profile aligned with the Tisza party, ChatGPT did not recommend it. In percentage-matching assessments, Tisza received a score in just 2% of the responses. In contrast, views aligned with Fidesz were identified much more consistently, while 96% of responses from both ChatGPT and Gemini included at least one party not present on the 2026 ballot.

      The researchers found no indication that this disparity was intentionally created, nor did they assert that chatbot responses influenced the election outcome. Potential factors like training-data gaps, safety filters, language-processing limitations, and Tisza's rapid ascent post-2024 may have played a role. However, these explanations provide little reassurance to voters receiving well-presented and compelling recommendations from systems with opaque reasoning.

      This cautionary note is part of a growing body of similar research. During Scotland's 2026 election, Demos evaluated five AI services with 75 election-related inquiries and discovered factual inaccuracies in 34.1% of responses. ChatGPT made errors in 46.2% of cases, including incorrect election dates and eligibility criteria, fictitious candidates, and invented political controversies. Nearly half of the responses lacked citations or supporting links.

      In 2025, a Dutch regulator reached another concerning finding. Despite the country's diverse multiparty system, four tested chatbots directed voters to only two major parties in 56% of interactions. Additional academic research has identified recurring patterns of political preference among ChatGPT and Gemini, although the nature and intensity of these patterns can vary based on models, prompts, languages, and elections.

      Politicians are already discovering ways to influence chatbot responses.

      Accuracy and inherent bias are just part of the issue. The information feeding these systems can also be deliberately manipulated. The New York Times recently highlighted Missouri political candidate Dustin Lloyd, whose priorities were hardly reflected in chatbot responses. Lloyd created a meticulously curated question-and-answer page on his campaign website. Following this, chatbot responses began to link his personal background with his policy objectives, demonstrating how quickly a campaign can shape the AI-generated narrative of its candidate.

      Keeping an accurate campaign website is not inherently malicious. However, this same mechanism presents a clear opportunity for exaggerated claims, attack pages, fake organizations, and sites primarily aimed at influencing AI responses.

      A BBC investigation highlighted how low the threshold for technical manipulation can be. A journalist took about 20 minutes to publish a fake blog post claiming to be the world's best hot-dog-eating technology reporter. Within a day, ChatGPT, Gemini, and Google’s AI Overviews were regurgitating parts of the fabricated tale, while Claude managed to resist the trap.

      Likewise, another BBC report examined Google's efforts to counter this emerging manipulation trend. The company has documented websites designed to hijack AI systems, sway recommendations, promote specific businesses, and potentially steal information. Google anticipates that these indirect prompt-injection tactics will become more widespread and sophisticated.

      Researchers have also demonstrated that AI-enhanced search engines remain vulnerable to specialized manipulation techniques. Approaches such as rewritten-query stuffing and splitting promotional text into segments significantly increased the manipulation rate compared to a baseline attack in one 2026 study. In response, Google has expanded its spam regulations to encompass tactics aimed at distorting answers produced by AI Overviews and AI Mode.

      An electoral bomb with an unpredictable impact.

      AI chatbots possess several unstable characteristics. Their responses can be inaccurate, politically biased, and persuasive. Additionally, those answers are challenging to reproduce and reliant on websites that can be altered by campaigns or malicious actors. The risk goes beyond a chatbot explicitly endorsing an incorrect candidate.

      Potential electoral damage could stem from omitted parties, outdated voting information, a fabricated scandal, or a strategically placed webpage that gets integrated into an authoritative-sounding response. AI has essentially become an election bomb, with politicians, platforms, researchers, and opportunistic manipulators all having their hands on the trigger—and the potential fallout remains uncertain.