A Liberties report testing ChatGPT and Google’s Gemini during Hungary’s parliamentary election found the chatbots gave inaccurate, inconsistent and unreproducible voting advice. Researchers created five voter profiles aligned to the five national-list parties defined by Voksmonitor and ran each profile 10 times in ChatGPT and 10 times in Gemini, using direct advice and percentage-match prompts. ChatGPT assigned the winning opposition Tisza party a match score in just 2 percent of percentage tests and failed to recommend Tisza in 90 percent of cases when fed Tisza-aligned profiles. By contrast, Fidesz-aligned profiles were identified far more consistently; in direct-advice prompts ChatGPT named Fidesz as the single recommended party in about 50 percent of trials. Responses were highly volatile, with materially different answers to identical prompts, and 96 percent of replies from ChatGPT and Gemini included parties not on the 2026 national ballot. The report notes these general-purpose AI systems do not disclose how they generate political guidance, do not yield reproducible results, and are not subject to election-related oversight, and warns that in closer races or among uncertain voters such outputs could potentially affect outcomes.
