A study released on July 21, 2026 by Liberties, a civil liberties advocacy group, reveals that ChatGPT and Gemini provided inaccurate, inconsistent, and unreliable voting advice during Hungary's 2026 parliamentary elections. In 96% of cases, both chatbots cited parties absent from Hungary's national ballot. ChatGPT failed to recommend Tisza—the election winner—in 90% of cases when tested with voter profiles aligned to that party. The study identifies a regulatory gap in European law and recommends halting personalized AI-driven voting recommendations.
ChatGPT failed to recommend Tisza in 90% of test cases with aligned voter profiles
96% of responses cited parties not on Hungary's 2026 national ballot
Both chatbots displayed disclaimers before delivering partisan recommendations
A regulatory gap identified between the EU's AI Act and Digital Services Act
Liberties urges halt to personalized voting recommendations by AI systems
A study published on July 21, 2026 by Liberties, a civil liberties advocacy group, concludes that general-purpose AI chatbots provide inaccurate, inconsistent, and unreliable voting advice. These findings are based on tests conducted during Hungary's 2026 parliamentary elections.
Liberties selected ChatGPT and Gemini for testing—the two most popular AI systems among the 29.8% of Hungary's population identified as AI users. The methodology involved five distinct voter profiles, each submitted ten times to each model with two different types of queries.
These profiles were built from official party positions listed in Voksmonitor, a Hungarian voting advice application. Each profile aligned with one of the five parties registered on Hungary's national ballot.
Systematic biases in the responses
In 96% of cases, ChatGPT and Gemini cited parties not appearing on Hungary's 2026 national ballot.
Reliability of ChatGPT and Gemini voting advice in Hungary's 2026 elections (% of test cases — source: Liberties study)
The Tisza Party (Respect and Liberty Party, a centre-right formation founded in 2020 and led by Péter Magyar) decisively won Hungary's 2026 parliamentary elections. It ended 16 years of rule by Viktor Orbán and his Fidesz party.
Yet when ChatGPT received voter profiles aligned with Tisza, it failed to recommend that party in 90% of cases. In the percentage-matching tests, the model assigned Tisza a score in just 2% of responses.
ChatGPT often directed these voters toward a small party below the 5% threshold or toward parties absent from the national ballot. For voter profiles aligned with Fidesz, the model identified that party as the sole recommendation in approximately 50% of cases, and presented it as the primary option in the remaining responses.
Disclaimers followed by partisan recommendations
The study identifies a paradoxical behaviour common to both models. ChatGPT and Gemini often began their responses with disclaimers stating they cannot provide political advice. They then proceeded with dense paragraphs of persuasive partisan recommendations.
cannot provide political advice
— Disclaimer regularly displayed by ChatGPT and Gemini at the start of their responses, before delivering partisan recommendations (Liberties study, July 21, 2026)
Context: Training data failing to capture Tisza's rise
The probable cause of the errors lies in training data gaps. Since the Tisza Party only emerged as a major political force after 2024, static AI models prove unable to correctly position it within Hungary's political landscape.
Timeline
2010
Viktor Orbán and his Fidesz party come to power in Hungary, beginning 16 years of political dominance.
2020
The Tisza Party (Respect and Liberty Party) is founded by Péter Magyar in Hungary.
October 19, 2022
The European Union adopts the Digital Services Act (EU Regulation 2022/2065).
After 2024
Tisza emerges as a major political force in Hungary—too recently to be well represented in the training data of large AI models.
2026 (elections)
Tisza decisively wins Hungary's parliamentary elections, ending 16 years of Viktor Orbán's rule.
July 21, 2026
Liberties releases its study on the reliability of AI chatbots for voting advice in Hungary's elections.
The European regulatory gap
The study identifies a structural regulatory gap. The EU's AI Act requires providers of general-purpose AI models (GPAI) to assess potential systemic risks.
The Digital Services Act (EU Regulation 2022/2065 on a single market for digital services, adopted October 19, 2022 to combat the spread of illegal content) covers systemic risks to electoral processes. Yet general-purpose AI chatbots fall between these two frameworks without being clearly covered by either.
Liberties' recommendations
Liberties recommends that AI providers stop offering personalized voting recommendations until they can guarantee transparency, accuracy, consistency, and accountability. The organization also calls for the development of safeguards for AI systems providing political advice.
What remains uncertain
The study characterizes the training data gap as the 'probable' cause but does not definitively establish it. Reactions from OpenAI and Google to the study's findings are not available from the sources. The study also does not clarify whether its results apply to other countries or elections.
FAQ
Why did AI chatbots get Hungarian political parties wrong?
The Liberties study identifies training data gaps as the probable cause. Since Tisza only emerged as a major force after 2024, AI models lacked sufficient information to correctly position it in Hungary's political landscape.
What is the Tisza Party?
Tisza (Respect and Liberty Party) is a centre-right Hungarian political party founded in 2020 and led by Péter Magyar. It won Hungary's 2026 parliamentary elections, ending 16 years of rule by Viktor Orbán and Fidesz.
Why were ChatGPT and Gemini chosen for testing?
These two systems are the most widely used among the 29.8% of Hungary's population identified as AI users.
Is there European regulation on AI voting advice?
The study identifies a regulatory gap. The AI Act imposes obligations on general-purpose AI model providers, and the Digital Services Act covers systemic risks to electoral processes—but AI chatbots fall between these two frameworks without being clearly subject to either.
What does Liberties recommend to AI providers?
Liberties recommends ending personalized voting recommendations until transparency, accuracy, consistency, and accountability can be guaranteed. The organization also calls for specific safeguards for systems providing political advice.