A study conducted by the civil liberties organization Liberties during this year’s Hungarian parliamentary elections has found that popular AI chatbots provide voters with significantly flawed and inconsistent guidance on which political party to support. The research tested ChatGPT and Gemini, the two most widely used AI systems in Hungary, by creating five voter profiles aligned with different parties and querying each system multiple times with identical prompts seeking voting advice.
The study identified substantial problems with the chatbots’ outputs. ChatGPT failed to recommend Tisza, the opposition party that won decisively under leader Péter Magyar, in 90% of cases when presented with voter profiles aligned with that party’s positions. When tested on percentage-matching queries, ChatGPT assigned Tisza a score in only 2% of responses. In contrast, the previously governing Fidesz party was consistently identified by ChatGPT as a primary voting option in about half of direct advice queries. Additionally, both systems included parties not appearing on the 2026 ballot in 96% of their responses.
A critical concern raised by researchers is the apparent reliability of the chatbots’ responses despite their inaccuracy. Both systems began interactions with disclaimers stating they could not provide political advice, yet proceeded to deliver detailed party recommendations presented as authoritative and well-reasoned. This presentation style creates a risk that users will trust the outputs despite the underlying methodology being opaque and results being unstable. Identical voter profiles yielded materially different recommendations in successive tests.
The researchers attributed the failures partly to training data limitations, noting that Tisza’s rise to prominence occurred after 2024, making it difficult for static AI models to accurately position the party. However, they emphasized a broader regulatory gap: the EU’s AI Act and Digital Services Act do not adequately cover AI chatbots providing political matching services. The study calls for safeguards requiring transparency, accuracy, consistency and accountability in any AI system offering personalized voting guidance.
Article Attribution | Read More at Article Source
Article summary produced by Claude AI