AI Chatbots Give Unreliable Voting Advice, Hungary Study Finds

Major AI chatbots delivered unreliable and often inaccurate voting guidance during Hungary's 2026 parliamentary elections, according to new research that raises questions about artificial intelligence's role in democratic processes.

A study by civil liberties organization Liberties found that ChatGPT and Gemini frequently recommended parties not running in the election, omitted relevant parties from their suggestions, and produced wildly inconsistent results when given identical voter profiles. The findings were first reported by The Guardian.

Systematic visibility gaps

The most striking discovery involved Tisza, the opposition party that won the election in a landslide victory ending Viktor Orbán's 16-year tenure as prime minister. When researchers fed ChatGPT detailed voter profiles aligned with Tisza's positions, the AI failed to recommend the party in 90% of cases.

In percentage-matching tests, ChatGPT assigned Tisza a compatibility score in just 2% of attempts. Instead, voters with Tisza-aligned views were routinely steered toward smaller parties unlikely to meet the 5% parliamentary threshold or parties not even on the national ballot.

By contrast, profiles matching Fidesz—Orbán's national-conservative party—were recognized far more consistently. ChatGPT identified Fidesz as the recommended party in roughly 50% of direct advice prompts and presented it as a primary option in most remaining cases.

Testing methodology

Liberties researchers created five distinct voter profiles based on party positions from Voksmonitor, a respected Hungarian voting advice application. Each profile aligned with one of the five parties registered for the parliamentary elections.

The team tested each profile 10 times in both ChatGPT and Gemini using two different prompt types: requests for direct party recommendations and requests for percentage matches with each party. The results revealed profound inconsistencies—the same voter profile often matched with radically different parties in successive tests.

Both AI models included parties not on the 2026 ballot in 96% of their responses. The systems also regularly began with disclaimers stating they "cannot give political advice" before delivering several paragraphs of detailed party recommendations that appeared authoritative and well-reasoned.

Why it matters

With nearly 30% of Hungary's population identified as AI users, the study exposes a regulatory blind spot at the intersection of artificial intelligence and electoral integrity. While the research doesn't claim the AI guidance affected the election outcome—Tisza won decisively—the findings suggest potential risks in closer contests or among less certain voters.

The underlying causes likely stem from training data limitations. Tisza surged to prominence only after 2024, meaning static AI training models lacked sufficient information to accurately represent the party. But the broader issue involves transparency: general-purpose AI systems don't disclose how they generate political guidance, don't produce reproducible results, and aren't subject to election-related oversight.

"General-purpose AIs should not present opaque and unstable political matching as if it were reliable electoral guidance," said Eva Simon, head of Liberties's tech and rights programme. "Democracy cannot rely on opaque systems that claim neutrality while delivering advice they cannot explain, reproduce or guarantee to be accurate."

The report argues that AI chatbots currently fall between regulatory frameworks—the EU's AI Act covers general-purpose AI model risks, and the Digital Services Act addresses electoral process risks, but chatbots providing voting advice aren't clearly covered by either.

Liberties recommends that AI providers either stop offering personalized voting recommendations or implement safeguards guaranteeing transparency, accuracy, consistency, and accountability. The research was first reported by The Guardian.