Key facts
- AI chatbots like ChatGPT and Gemini provided inaccurate and unreliable voting advice in a study.
- The chatbots often failed to recommend the correct party or included parties not running in elections.
- In 90% of cases, ChatGPT failed to recommend the Tisza party when fed a Tisza-aligned voter profile.
- ChatGPT assigned Tisza a score in only 2% of percentage-matching tests.
- Fidesz was recommended more consistently by ChatGPT for aligned voter profiles.
- The study found AI responses were volatile, with materially different answers to identical prompts.
- AI systems did not disclose their methods for generating political guidance and lacked public oversight.
AI chatbots provide inaccurate, inconsistent, and unreliable guidance to voters regarding which party to support, according to a study conducted during Hungary's parliamentary elections. The research, published by the civil liberties group Liberties, found that systems like ChatGPT and Gemini often failed to recommend the correct party, omitted relevant ones, or listed parties not even participating in the election.
In tests based on a respected Hungarian voting advice app, researchers created five distinct voter profiles. When fed a detailed profile aligned with the opposition Tisza party, ChatGPT failed to recommend it in 90% of cases and assigned it a score of just 2% in percentage-matching tests. Conversely, Fidesz-aligned profiles were recognized more consistently, with ChatGPT identifying Fidesz as the primary voting option in about 50% of direct advice prompts.
The study highlighted that unlike dedicated voting advice sites, general-purpose AI systems do not disclose their methods, lack reproducibility, and are not subject to electoral oversight, despite often appearing confident and authoritative. This opacity and instability pose a risk to voters, particularly in competitive elections.
Researchers suggested that training data gaps, especially for parties that surged in prominence recently like Tisza, likely contribute to these errors. The findings also point to a regulatory gap, as current EU legislation may not fully cover the risks posed by AI chatbots in electoral contexts. Liberties urged AI providers to stop offering personalized voting recommendations unless they can guarantee transparency, accuracy, consistency, and accountability.