When voters asked ChatGPT which party to back in Hungary’s 2026 parliamentary election, the chatbot routinely recommended parties that weren’t even on the ballot. When fed a voter profile aligned with the party that went on to win in a landslide, it failed to recommend that party 90% of the time. And when the same question was asked twice, the answers were often materially different. These are the findings of a study published by civil liberties group Liberties, based on research during Hungary’s April election and reported by The Guardian.
What is this study? Liberties, a European civil liberties organisation, created five distinct voter profiles aligned with each of the five parties registered on Hungary’s national ballot lists. Each profile was tested 10 times in ChatGPT and 10 times in Gemini, using two types of requests: direct advice on which party to back, and a percentage match with each party. The profiles were based on party positions outlined on Voksmonitor, a respected Hungarian voting advice application. The study did not test whether AI changed the election outcome — Tisza won decisively — but it did test whether AI could reliably tell a voter who to vote for.
🔍 THE BOTTOM LINE
The two most popular AI chatbots in Hungary gave voting advice that was wrong, inconsistent, and biased against the party that actually won. In 96% of responses, they included parties not on the ballot. In 90% of cases where a voter’s profile aligned with the winning Tisza party, ChatGPT failed to recommend it. The chatbots sounded confident. The answers were not reproducible. This is not a tool you want between a voter and a ballot box.
What the Data Shows
The numbers are stark. In percentage-matching tests, ChatGPT assigned the eventual winner — Péter Magyar’s opposition Tisza party — a score in just 2% of cases. Instead, Tisza-aligned voters were routinely steered toward a smaller party unlikely to cross the 5% parliamentary threshold, or toward parties not on the national ballot at all.
By contrast, Fidesz-aligned voter profiles were recognised far more consistently. In direct advice prompts, ChatGPT identified Fidesz — Viktor Orbán’s national-conservative party that had governed Hungary for 16 years — as the single party the voter should back in about 50% of cases, presenting it as a primary option in the remainder.
In 96% of responses from both ChatGPT and Gemini, the chatbots included parties that were not on the 2026 ballot. The same voter profile, asked the same question twice, could receive radically different party recommendations. The study described these as “materially different” answers to identical prompts.
Why This Matters Beyond Hungary
Hungary’s election was not close — Tisza won decisively, ending Viktor Orbán’s 16-year grip on power. The study is careful to note it does not claim the outcome was affected. But the researchers warn: “In a more competitive election or where voters are less certain, such outputs could potentially impact the outcome.”
That warning applies to every democracy. The Reuters Institute’s Digital News Report 2026 found that 10% of news consumers globally use AI chatbots to get their news, rising to 16% for under-35s. If one in ten voters is getting political guidance from a system that invents parties, misses the winner, and changes its answer when asked twice, the democratic risk is not hypothetical.
The core problem identified by the researchers is one of authority without accountability. Unlike voting advice applications (which disclose their methodology, produce reproducible results, and are subject to election oversight), general-purpose AI systems “did not disclose how they generated political guidance, did not yield reproducible results, and were not subject to election-related public oversight.” Their responses look confident and authoritative, making users more likely to trust them.
The NZ Angle
New Zealand uses MMP — a system where small parties can cross thresholds and coalition negotiations are complex. A chatbot that routinely recommends parties below the threshold, or omits the party a voter is most aligned with, is not a minor bug in an MMP context. It’s a structural distortion of the voter’s decision.
NZ also has no specific regulation of AI in electoral contexts. The Electoral Commission provides official voting guidance, and independent platforms like Vote Compass offer transparent methodology. But nothing stops a voter from asking ChatGPT “which party should I vote for?” — and nothing requires the chatbot to be right, transparent, or even consistent.
The Hungarian study suggests it will be none of those things. For NZ voters who use AI tools — and that proportion is growing — the risk is the same as in Hungary: confident-sounding advice that omits the right answer, invents non-existent options, and changes when asked twice.
What “Unreliable” Actually Means Here
This is not a case of AI being slightly off. The failure modes are specific and severe:
- Omission of the correct party — in 90% of Tisza-aligned tests, ChatGPT did not recommend Tisza.
- Invention of non-existent parties — in 96% of responses, chatbots listed parties not on the ballot.
- Volatility — identical prompts produced materially different recommendations across runs.
- Asymmetric bias — Fidesz-aligned profiles were recognised consistently; Tisza-aligned profiles were not.
The asymmetry is the most concerning finding. The chatbot was not randomly wrong — it was systematically less likely to recommend the opposition party that won the election. Whether this reflects training data bias, index bias, or something else, the study doesn’t determine. But the pattern is clear: the AI was not a neutral adviser.
❓ FAQ
Did AI change the Hungarian election result? No. The study explicitly says it does not claim the outcome was affected — Tisza won decisively. The concern is what happens in a closer election where voters are less certain.
Which chatbots were tested? ChatGPT and Gemini, the two most popular AI tools among Hungary’s 29.8% AI-using population.
Why did the chatbots recommend parties not on the ballot? The study doesn’t determine the cause, but it’s likely a combination of outdated training data, hallucination, and the chatbots’ general-purpose design — they’re not built as electoral tools and don’t have real-time ballot access.
Does NZ have rules on AI in elections? Not specifically. The Electoral Commission provides official guidance, and platforms like Vote Compass offer transparent methodology, but there’s no regulation preventing voters from relying on AI chatbots for voting advice.
What’s the fix? The researchers argue general-purpose AI systems should not be used for voting guidance without disclosure of methodology, reproducible results, and election oversight — the standards that dedicated voting advice applications already meet.
🔍 THE BOTTOM LINE
A chatbot that invents parties, omits the winner, and changes its answer when you ask twice is not a voting advisor. It’s a slot machine with confident typography. The Hungarian study is the first detailed field test of AI voting advice in a real election, and the results are unambiguous: the two most popular chatbots failed on every measure that matters — accuracy, consistency, and neutrality. Every democracy with an election coming up — including NZ’s — should treat this as a warning, not a curiosity.
