During the 2026 Hungarian election campaign, researchers at Liberties asked the free versions of ChatGPT and Gemini the question a lot of voters are quietly asking their phones: based on what I believe, who should I vote for?
Across 200 outputs, with voter profiles built from a real voting advice application, both systems failed under conditions far easier than real life. ChatGPT identified the governing party's profile correctly half the time. For the profile matching Tisza, the opposition party that went on to win with a two thirds majority, it got it right zero times out of ten.
Both systems also opened by saying they could not tell anyone how to vote, then gave detailed, confident advice anyway. The disclaimer did not work as a safeguard. If anything, it made the answer that followed look more credible than it deserved to.
Eva Simon leads the Tech and Rights Program at Liberties and co-authored the report. She will join us via video to walk us through what the outputs actually looked like, why the failures are structural rather than a bug someone can patch, and where the AI Act and the DSA currently leave this.
The interesting question is not only what the research found, but what we do with it.
The event is hosted by Demai, an impact business building technologies like Parl8 (www.parl8.eu) to develop and defend democracy. Holding events that share knowledge from where democracy and technology meet is part of that mission.
Free, in English, everyone welcome. Report: https://www.liberties.eu/en/stories/hungary-elections-ai/45665
