Election voting advice from AI chatbots ‘inaccurate and unreliable’

AI chatbots ChatGPT and Gemini provided inaccurate and inconsistent voting guidance during Hungary's April 2024 parliamentary elections, according to research by civil liberties group Liberties. Testing five voter profiles against the country's five registered parties, ChatGPT assigned the opposition Tisza party a score in just 2% of percentage-matching tests, despite Tisza winning the election decisively. The systems failed to disclose how they generated recommendations, yielded inconsistent results to identical prompts, and lacked public electoral oversight, raising concerns about their reliability in electoral contexts.

The study tested ChatGPT and Gemini on 29.8% of Hungary's population identified as AI users, using profiles based on party positions from the Voksmonitor voting advice app. ChatGPT consistently recommended smaller parties or non-registered parties to Tisza-aligned voters while reliably identifying Fidesz as a primary option for Fidesz-aligned profiles, suggesting algorithmic bias in party visibility. Researchers cautioned that while Tisza's landslide victory over Prime Minister Viktor Orbán's 16-year reign appeared unaffected, similar outputs could impact closer elections or less-certain voters.