Election voting advice from AI chatbots ‘inaccurate and unreliable’

Research during Hungary election shows AI recommended parties not running and gave highly volatile answers to identical prompts

AI chatbots provide inaccurate, inconsistent and unreliable guidance to voters asking which party they should back, a study suggests, often recommending the wrong party, not mentioning the right one, or listing parties not even running.

“The results raise serious concerns about the reliability of general-purpose AI systems in electoral contexts,” the study , published by the civil liberties group Liberties and based on research during this year’s Hungarian parliamentary elections, concluded.

“They misclassified profiles, omitted relevant parties and included parties not running in the election,” said the report, shared with the Guardian. Outcomes were also highly volatile, with “materially different” answers given to identical prompts.

The report’s most striking finding from the Hungarian election – convincingly won by Péter Magyar ’s opposition Tisza party – was an overwhelming bias in party visibility, particularly as regards Tisza: in 90% of cases when ChatGPT was fed a detailed Tisza-aligned voter profile, it failed to recommend the party.

Magyar’s landslide brought to an end Viktor Orbán’s 16-year grip on power as prime minister and head of his national-conservative Fidesz party.

The researchers noted that unlike voting advice sites or independent media explainers, general-purpose AI (GPAI) systems did not disclose how they generated political guidance, did not yield reproducible results, and were not subject to election-related public oversight.

However, their responses often look confident and authoritative, making users more likely to trust them. With 29.8% of Hungary’s population identified as AI users, the researchers tested the advice given by the two most popular, ChatGPT and Gemini.

Based on party positions outlined on Voksmonitor, a respected Hungarian voting advice app, Liberties created five distinct voter profiles, each aligned with one of the five parties registered on national lists for Hungary’s parliamentary elections.

Each voter profile was then tested 10 times in ChatGPT and 10 times in Gemini, using two different requests: one for direct advice on which party the fictitious voter should back, and another for a percentage match with each of the five parties.

In percentage-matching tests, ChatGPT assigned Tisza a score in just 2% of cases. Instead, users with Tisza-aligned views were routinely recommended a smaller party unlikely to cross the 5% threshold for parliament, or parties not on the national ballot.

By contrast, Fidesz-aligned voter profiles were recognised far more consistently. In direct advice prompts, ChatGPT identified Fidesz as the single party the user should vote for in about 50% of cases, presenting it as a primary option in the remainder.

The analysis does not claim that the outcome of the election was affected: Tisza still won decisively. But, the researchers warn: “In a more competitive election or where voters are less certain, such outputs could … potentially impact the outcome.”

Other issues included inconsistency, with the same voter profile matched with radically different parties in successive tests, and the inclusion – in 96% of responses from ChatGPT and Gemini – of parties not on the 2026 ballot.

The researchers also noted that both AI models regularly began their responses with polite disclaimers saying they “cannot give political advice” – before providing several paragraphs of dense, highly persuasive party recommendations.

“The answers appeared well-argued, precise and authoritative,” the researchers said. “This creates a risk that users may treat the outputs as reliable, even though the underlying method is opaque and the results are unstable.”

The report argues that the underlying cause of the errors probably lies in training data gaps, filters and language-processing limitations. Tisza surged to prominence only after 2024, meaning static AI training models struggled to place it on their map.

But they said the findings exposed a serious regulatory gap: the EU’s AI Act obliges providers of GPAI models to assess systemic risks, and its Digital Services Act covers “systemic risks” to electoral processes – but AI chatbots fall between the two.

Liberties said safeguards should be developed for AI systems providing political advice, and AI providers should stop offering personalised voting recommendations unless they could guarantee transparency, accuracy, consistency and accountability.

“General-purpose AIs should not present opaque and unstable political matching as if it were reliable electoral guidance” for voters, said Eva Simon, the 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.”