A Liberties study of AI voting advice during Hungary’s parliamentary election found ChatGPT and Gemini gave inaccurate, inconsistent and unreproducible guidance. Researchers created five voter profiles based on Voksmonitor and ran each profile 10 times in ChatGPT and 10 times in Gemini, using two prompt types: direct party advice and percentage matches. When fed detailed profiles aligned with Péter Magyar’s Tisza party, ChatGPT failed to recommend Tisza in 90 percent of direct-advice cases and assigned Tisza a score in only 2 percent of percentage-matching tests. By contrast, Fidesz-aligned profiles were identified far more consistently, with ChatGPT naming Fidesz as the single party to back in about 50 percent of direct-advice prompts. Responses were highly volatile, producing materially different party recommendations to identical prompts, and in 96 percent of responses the chatbots listed parties that were not on the 2026 national ballot. The report also noted that general-purpose AI systems do not disclose how they generate political guidance, do not produce reproducible outputs, and are not subject to election-specific oversight, and warned that in closer contests or with less certain voters such outputs could potentially affect outcomes.
