AI Chatbots Are Giving Voters Wrong Answers About Which Party to Support

Key points

• A study published by civil liberties group Liberties found AI chatbots gave inaccurate, inconsistent voting advice during Hungary's 2025 parliamentary elections.
• Chatbots sometimes recommended political parties that were not running in the election at all.
• Identical questions put to the same AI tools produced wildly different answers each time.
• Researchers concluded that general-purpose AI systems cannot be relied on for electoral guidance.

Ask a chatbot which party you should vote for and you might get an answer. A confident, well-written answer. But a new study suggests it could be completely wrong.

Researchers tested AI chatbots, conversational tools powered by large language models (the same underlying technology that runs ChatGPT), during Hungary's parliamentary elections earlier this year. The tools recommended parties not standing in the election, skipped parties that were on the ballot, and gave starkly different answers when asked the exact same question twice. Like asking a friend for directions and getting a different route every time, including one that sends you to a street that doesn't exist.

The study was published by Liberties, a European civil liberties organisation, and first reported by The Guardian. Real prompts, real election, real mess.

Should voters be worried?

Yes, if they plan to use a chatbot as a shortcut to deciding how to vote. The researchers described the results as raising "serious concerns" about using general-purpose AI in electoral settings. These tools are built to be helpful across thousands of topics. Elections require precision: the right parties, the right dates, the right rules. That precision is exactly where the chatbots stumbled.

Practically speaking, if you ask an AI assistant something like "which party best matches my views on healthcare?", the answer may name a party that folded two years ago, miss the obvious frontrunner, or give your neighbour a completely different list if they type the same words an hour later.

The inconsistency matters as much as the inaccuracy. A tool that's sometimes wrong is hard to trust. A tool that's randomly wrong is worse, because you can't even know when to double-check.

This isn't an argument that AI tools are useless. For drafting a meal plan or summarising a long document, they do fine. Voting sits in a different category: the stakes are civic, the facts are specific, and a hallucination (when an AI confidently states something false) can nudge someone toward a party that doesn't exist. Our earlier reporting on AI agents breaking things for real customers after passing internal tests shows this gap between lab performance and real-world reliability isn't unique to elections.

The practical advice is simple. Use official government election websites or your country's electoral commission for any voting information. Treat chatbot answers about elections the way you'd treat directions from a stranger who's never visited your town: possibly interesting, definitely unverified.