AI chatbots gave unreliable voting advice in Hungary study

AI chatbots gave inaccurate, inconsistent and unreliable guidance to voters seeking party recommendations during Hungary’s parliamentary elections, according to research by the civil liberties group Liberties. Researchers tested ChatGPT and Gemini with voter profiles aligned to parties registered on national lists, finding that the systems misclassified preferences, omitted relevant parties and sometimes recommended parties that were not running.

The most striking visibility gap involved Péter Magyar’s opposition Tisza party, which won the election and ended Viktor Orbán’s 16-year hold on power. ChatGPT failed to recommend Tisza in 90% of cases when given a detailed Tisza-aligned voter profile, while percentage-matching tests assigned Tisza a score in just 2% of cases. Tisza-aligned users were often directed toward a smaller party unlikely to cross the 5% threshold for parliament or parties absent from the national ballot.

The systems also produced materially different answers to identical prompts and included parties not on the 2026 ballot in 96% of responses from ChatGPT and Gemini. Liberties said general-purpose AI tools lack the transparency, reproducibility and electoral oversight associated with voting advice sites, and urged providers to stop offering personalised voting recommendations unless they can guarantee transparency, accuracy, consistency and accountability.