Another election cycle, another stark warning from researchers: turning to AI chatbots for political guidance is a profoundly ill-advised endeavor. Recent analyses conducted during Hungary’s 2026 parliamentary elections have definitively concluded that prominent AI models, including ChatGPT and Google Gemini, are not only providing inaccurate but also alarmingly inconsistent and unreliable voting advice. These sophisticated algorithms have demonstrated a troubling propensity to misclassify voter profiles, overlook significant political parties, recommend candidates absent from the official ballot, and even generate materially different responses when presented with identical information multiple times. This pattern of failure is not isolated, raising serious concerns about the integrity of information voters receive as AI becomes increasingly integrated into online search and information retrieval.
The Persistent Failure of AI in Voter Assessments
The implications of these findings are significant, particularly as AI tools become more sophisticated and widely adopted by the public. The Civil Liberties Union for Europe, in a comprehensive study designed to test the political neutrality and accuracy of AI chatbots, created five distinct fictional voter profiles. These profiles were meticulously crafted to align with the stated positions of the five major parties contesting Hungary’s national elections. Each profile was then subjected to repeated questioning, with prompts designed to elicit direct voting recommendations and percentage-based party alignments.
The results were deeply concerning. ChatGPT, for instance, failed to recommend the opposition Tisza party in a staggering 90% of tests where the fictional voter profile was clearly aligned with Tisza’s platform. During percentage-matching exercises, where users sought to quantify their political leanings, ChatGPT assigned Tisza a score in a mere 2% of its responses. Conversely, views aligned with the Fidesz party were identified with considerably more consistency. A similarly worrying trend emerged when the chatbots were asked to identify relevant parties: an astounding 96% of responses from both ChatGPT and Gemini included at least one political party that was not even on the official 2026 Hungarian ballot.
While the researchers found no direct evidence of deliberate manipulation or bias engineered into the AI models, they acknowledged that several factors could contribute to these skewed outcomes. These include potential gaps in the training data, the implementation of safety filters designed to prevent harmful outputs, inherent limitations in natural language processing capabilities, and the rapid rise of the Tisza party after 2024, which may not have been fully reflected in older training datasets. Nevertheless, these explanations offer little reassurance to voters who are increasingly relying on the seemingly authoritative and polished recommendations generated by these opaque systems.
A Pattern of Errors and Omissions in AI-Generated Political Information
This Hungarian study is not an outlier; it is the latest in a growing body of research highlighting the unreliability of AI in political contexts. A similar investigation conducted during Scotland’s 2026 elections by Demos tested five different AI services with 75 election-related questions. Their findings revealed factual inaccuracies in a significant 34.1% of the responses. ChatGPT, in particular, exhibited a higher error rate, producing incorrect information in 46.2% of its responses. These errors ranged from misstating election dates and eligibility rules to fabricating candidates and even inventing political scandals. Furthermore, a disturbing nearly half of the responses provided no citation or supporting link, making it impossible for users to verify the information presented.
In 2025, a Dutch regulatory body also arrived at a disquieting conclusion. Despite the Netherlands’ robust and diverse multiparty system, four tested AI chatbots consistently directed voters toward only two prominent parties in 56% of their interactions. These findings suggest a systemic tendency for AI models to oversimplify complex political landscapes and favor established or more frequently discussed political entities. Separate academic investigations have corroborated these observations, identifying recurring patterns of political preference across different AI models, including ChatGPT and Gemini. While the direction and severity of these biases can fluctuate depending on the specific model, the nature of the prompts, the language used, and the electoral context, the underlying issue of embedded bias remains a persistent concern.
The Evolving Landscape: Campaigns Learn to Shape AI Responses
Beyond the inherent inaccuracies and embedded biases, a more insidious challenge is emerging: the deliberate manipulation of the information that feeds these AI systems. Political campaigns are beginning to understand how to shape AI-generated narratives to their advantage. A recent report by The New York Times detailed the case of Dustin Lloyd, a political candidate in Missouri, whose campaign priorities were poorly reflected when voters inquired about him through AI chatbots. In response, Lloyd’s campaign strategically published a detailed question-and-answer page on their website. Subsequently, chatbot responses began to more accurately connect his personal background with his policy objectives, illustrating the speed at which a campaign can influence the AI-generated portrayal of its candidate.
While maintaining an accurate and comprehensive campaign website is a standard and legitimate practice, the underlying mechanism presents a clear pathway for malicious actors. This same technique can be exploited to disseminate exaggerated claims, create fabricated attack pages, launch deceptive organizations, and curate websites specifically designed to influence AI responses. The ease with which such manipulation can occur raises serious questions about the authenticity of information voters might encounter.
A BBC investigation starkly demonstrated the low technical barrier to such manipulation. A journalist successfully fabricated a blog post within approximately 20 minutes, falsely declaring himself the world’s foremost hot-dog-eating technology reporter. Within 24 hours, major AI platforms, including ChatGPT, Gemini, and Google’s AI Overviews, began incorporating elements of this invented story into their responses, showcasing the rapid propagation of misinformation. Notably, Claude, another AI model, demonstrated greater resistance to this fabricated narrative.
Further underscoring these vulnerabilities, another BBC report explored Google’s efforts to combat this burgeoning industry of AI manipulation. The company has identified websites containing explicit instructions designed to hijack browsing AI systems, influence recommendations, promote specific businesses, and potentially facilitate data theft. Google anticipates that these indirect prompt-injection attempts will escalate in both scale and sophistication.
Researchers have also demonstrated that AI-enhanced search engines remain susceptible to specially developed manipulation techniques. One 2026 study revealed that strategies such as rewritten-query stuffing and the segmentation of promotional text into discrete parts could double the manipulation rate compared to baseline attacks. In response, Google has expanded its spam policies to encompass attempts to distort answers generated by its AI Overviews and AI Mode features, indicating a growing awareness of the threat.
An Election Bomb with an Unpredictable Blast Radius
The confluence of these AI characteristics – inaccuracy, political unevenness, persuasive outputs, difficulty in reproduction, and susceptibility to external modification – creates a volatile and unpredictable scenario. The potential for electoral damage extends far beyond a simple case of a chatbot endorsing the wrong candidate.
The insidious nature of AI’s influence on elections could begin with seemingly minor issues: the omission of a relevant political party, outdated voting instructions, the dissemination of a fabricated scandal, or the strategic placement of a deceptive webpage that is then integrated into an authoritative-sounding AI response. AI has effectively become an election bomb. Politicians, tech platforms, researchers, and opportunistic manipulators all hold potential triggers, and the ultimate blast radius of this technological weapon remains largely unknown. The integrity of democratic processes hinges on the public’s ability to access accurate and unbiased information, and the current state of AI in political contexts poses a significant threat to this fundamental principle. As these technologies continue to evolve, a concerted effort from developers, regulators, and the public will be essential to mitigate the risks and ensure that AI serves as a tool for informed decision-making, rather than a vector for misinformation and manipulation.
