Researchers have published findings indicating that popular AI chatbots deliver unreliable electoral advice to voters, according to a study by civil liberties organization Liberties. The research, conducted during Hungary’s parliamentary elections earlier this year, examined how ChatGPT and Gemini responded to questions about which party voters should support based on their political preferences.
The analysis revealed multiple categories of errors in the systems’ outputs. Chatbots misidentified voter preferences, omitted relevant political parties from recommendations, and suggested parties that were not running in the election. Additionally, the systems produced highly inconsistent results when given identical voter profiles, demonstrating volatility in their recommendations. A particularly significant finding involved the Tisza party, which won the Hungarian election: when presented with voter profiles aligned with Tisza’s positions, ChatGPT failed to recommend the party in 90 percent of test cases.
Researchers constructed five distinct voter profiles based on positions from Voksmonitor, a Hungarian voting advice application, and tested each profile ten times with both ChatGPT and Gemini using two different request formats. The percentage-matching tests yielded especially problematic results, with ChatGPT assigning Tisza a score in only 2 percent of cases, while recommending smaller parties unlikely to achieve parliamentary representation. In contrast, profiles aligned with Fidesz received far more consistent recognition from the systems.
The researchers noted that general-purpose AI systems differ fundamentally from dedicated voting advice websites and independent media explainers in lacking transparency about how political recommendations are generated, failing to produce reproducible results, and remaining outside election-related regulatory oversight. Despite these limitations, the systems typically present their responses with apparent authority and confidence, potentially leading users to trust outputs that lack reliable methodological foundations.
Liberties identified a regulatory gap in current EU frameworks, noting that while the AI Act requires assessment of systemic risks from large language model providers and the Digital Services Act covers electoral risks, AI chatbots fall outside both regimes. The organization has called for the development of safeguards governing AI systems that provide political advice and recommends that providers cease offering personalized voting guidance unless they can guarantee transparency, accuracy, consistency and accountability.