If you want a quick recipe for lasagna or help debug JavaScript, chatbots work fine. Ask them who deserves your vote in an election, though, and you're stepping onto thin ice.
A study from the Civil Liberties Union for Europe (Liberties) examined how ChatGPT and Gemini handled political questions during the 2026 Hungarian parliamentary elections. The results were messy. Models frequently suggested political parties that weren't even on the ballot. They recommended tiny fringe parties that had zero chance of passing the 5% threshold, completely missed major opposition groups, and gave wildly different answers when handed the exact same prompt twice.
Yet, almost every response started with a polite disclaimer claiming the AI couldn't give political advice—right before firing off four paragraphs recommending specific parties.
That polite warning creates a false sense of security. People assume the system is giving neutral, well-balanced context. Instead, they get a lottery ticket wrapped in authoritative prose.
The Hungarian Case Study
To test how well these tools match voter values with actual platforms, researchers created five distinct voter profiles. They used party policy data from Voksmonitor, a respected voting advice app in Hungary. These profiles matched the platforms of five real, registered national parties running in the 2026 election, including Viktor Orbán's ruling Fidesz party and Péter Magyar's opposition Tisza party.
The outcome surprised even tech-skeptical researchers.
When fed a voter profile aligned with Tisza, ChatGPT failed to recommend the party 90% of the time. In percentage-matching tests, ChatGPT scored Tisza at just 2%. Instead, the AI regularly pointed Tisza-leaning voters toward tiny parties unlikely to cross the threshold, or parties that weren't even running in 2026.
Fidesz-aligned profiles didn't suffer from this invisibility problem. ChatGPT identified Fidesz as the direct choice about half the time and listed it as a top contender in almost every other test.
In total, 96% of the tested outputs from both OpenAI's ChatGPT and Google's Gemini included parties that weren't running on the 2026 national ballot.
Tisza went on to win the Hungarian election decisively anyway, so chatbot hallucinations didn't flip the outcome this time. But imagine a razor-thin contest where tens of thousands of voters rely on AI for a quick summary of who aligns with their values. Disappearing an entire main opposition group while boosting ghost candidates is a massive problem.
Why Chatbots Break Down on Ballots
General-purpose language models aren't built like traditional Voting Advice Applications (VAAs). Traditional voting tools follow transparent, auditable matrixes created by political scientists. They map precise policy stances against candidate platforms and show you exact math on how they reached a match.
AI chatbots don't work like that. They operate on probability, pattern recognition, and whatever training data they scraped months or years prior.
Several core flaws explain why political matching fails so hard in large language models.
Training Data Gaps and Recency Issues
Elections move fast. Parties split, form coalitions, adopt new manifestos, or launch completely new movements within months. Large language models struggle with recency. If a new political movement rises quickly—like Tisza did in Hungary—the model's underlying training data won't reflect its current strength, manifesto, or ballot status accurately.
The Illusion of Confidence
Chatbots are trained to sound helpful and authoritative. When a model lacks reliable data on local candidates or tactical voting options, it doesn't say "I don't know enough about this race." It fills the gaps with confident-sounding text, mixing real platform points with outdated or completely hallucinated details.
Opaque Logic and Zero Accountability
When an independent journalistic organization releases a voting guide, you can challenge their methodology. You can see who wrote it and evaluate their sources. General-purpose AI offers no such transparency. You can't audit why a model recommended a fringe party over a major opponent because the system itself cannot explain its internal weights.
What Needs to Change Before the Next Election Cycle
Nearly 30% of Hungary's population uses AI tools, and that number keeps rising across Europe and the US. Voters rely on these tools because modern election manifestos are long, dense, and painful to read.
Placing a soft disclaimer at the top of an AI response while providing flawed party rankings below it isn't enough. Here is what needs to happen to protect voters and keep political information accurate.
- Stop automated political matching: AI providers should disable direct party matching and voting recommendations until models can guarantee auditability and real-time accuracy.
- Mandate independent audits: Regulatory bodies and civil rights groups need access to standardized testing suites for AI systems during active election windows.
- Stick to official sources: If you're trying to figure out where candidates stand on key issues, skip the general chatbots entirely. Use accredited non-partisan voting guides, read candidate websites directly, or consult independent public media breakdowns.
- Treat AI outputs as raw draft text: Whenever an AI names a candidate or a party's stance, verify that the party is actually running on your local ballot before taking the advice seriously.
Don't outsource your civic duties to a probabilistic text generator. Checking primary sources takes five extra minutes, but it's the only way to make sure your vote actually goes where you want it to.
