Research conducted during Hungary’s parliamentary election found that leading AI chatbots repeatedly misidentified voters’ political preferences, recommended parties absent from the ballot and produced sharply different advice from identical requests.
Artificial intelligence chatbots may speak with the confidence of political experts, but new research suggests that voters should not rely on them to decide which party deserves their support.
A study by the Civil Liberties Union for Europe, known as Liberties, examined voting recommendations produced by OpenAI’s ChatGPT and Google’s Gemini during Hungary’s 2026 parliamentary election. Researchers concluded that both systems offered guidance that was frequently inaccurate, inconsistent and impossible for users to verify.
The investigation found that the chatbots misclassified political profiles, overlooked parties that closely matched the views presented to them and repeatedly included organisations that were not competing on Hungary’s national ballot. Identical prompts could also produce materially different answers when submitted more than once, undermining the idea that the systems could function as dependable voting-advice tools.
The most striking failure involved Tisza, the centre-right opposition movement led by Péter Magyar. When ChatGPT was given a detailed fictional voter profile aligned with Tisza’s policies, it failed to recommend the party in 90 percent of cases. During tests asking the chatbot to calculate compatibility percentages, Tisza received a score in only 2 percent of responses.
Instead, voters whose positions closely resembled Tisza’s programme were sometimes directed toward smaller parties with little chance of crossing Hungary’s 5 percent parliamentary threshold—or toward parties that were not standing in the election at all. Across the responses generated by ChatGPT and Gemini, parties absent from the 2026 ballot appeared in 96 percent of cases.
Fidesz, the national-conservative party led by Viktor Orbán, was recognised far more consistently. When presented with a Fidesz-aligned voter profile, ChatGPT identified it as the single party the user should support in approximately half of direct-advice tests and placed it among the principal choices in the remaining responses.
Researchers did not claim that the imbalance was deliberately programmed or that it changed the result of the election. Tisza ultimately won a decisive victory, ending Orbán’s 16 years in power and securing a comfortable parliamentary majority under Magyar.
The findings nevertheless raise questions about what could happen in a closer contest, particularly when voters are undecided or unfamiliar with the parties competing for office. A recommendation that excludes a major candidate, promotes an irrelevant organisation or changes each time the question is asked could distort a voter’s understanding of the available choices.
Liberties constructed five fictional voter profiles, each based on the positions of one of the five parties registered on Hungary’s national lists. The political preferences were derived from Voksmonitor, an established Hungarian voting-advice application. Each profile was submitted repeatedly to ChatGPT and Gemini using two formats: a direct request for a party recommendation and a request for percentage-based compatibility scores.
That methodology does not reproduce every way in which real voters interact with AI. It tested only two systems, used constructed profiles and did not measure whether chatbot responses actually persuaded anyone to change their vote. It did, however, demonstrate that apparently simple political-matching tasks could produce unstable outcomes even when the information supplied by the user remained unchanged.
The chatbots’ presentation made those failures potentially more consequential. Researchers found that the systems often began with disclaimers stating that they could not provide political advice, only to follow them with detailed and persuasive recommendations. The answers appeared structured, precise and authoritative despite the opacity of the process used to generate them.
Unlike conventional voting-advice applications, general-purpose chatbots do not normally explain how individual policy positions are weighted, which sources shaped the recommendation or why one party was ranked above another. Their answers are also probabilistic rather than fully reproducible, meaning the wording and conclusions can change between conversations.
One likely explanation for the poor treatment of Tisza is the party’s rapid rise. Magyar established the movement as a major national force only after 2024, leaving older training material with far more information about Fidesz and Hungary’s established political organisations. The report suggested that incomplete training data, model filters and difficulties processing political information in Hungarian may all have contributed.
The problem is not confined to Hungary. In the Netherlands, the national data-protection regulator warned voters before the October 2025 election that chatbots consistently failed as voting aids. Its tests found that the systems disproportionately directed users toward two large parties, even when the preferences entered into the chatbot reflected the programmes of smaller competitors.
A separate investigation ahead of Scotland’s 2026 parliamentary election found incorrect information in 34 percent of the answers supplied by five AI services. The systems invented candidates and scandals, misstated election dates and provided false information about voting requirements. An accompanying survey suggested that one in five voters had used AI chatbots or search tools for information about recent British elections.
Together, the studies point to a broader democratic risk. Chatbots are increasingly becoming substitutes for search engines, news websites and official information services. Their conversational style can make complex political issues easier to understand, but it can also conceal uncertainty behind fluent language and polished explanations.
Liberties argues that providers should stop offering personalised voting recommendations unless they can guarantee accuracy, consistency, transparency and accountability. The organisation also says the findings expose a regulatory gap between the European Union’s AI Act, which requires providers to assess certain systemic risks, and the Digital Services Act, which addresses threats to electoral processes on major online platforms.
AI systems can still help voters summarise manifestos, compare publicly stated policies and locate official electoral information. But the Hungarian study suggests that they should be treated as starting points for research rather than neutral political referees.
The central danger is not necessarily that a chatbot will openly instruct every user to support the same candidate. It is that unreliable systems may quietly make some parties more visible than others, while presenting unstable judgments as objective analysis.
In elections decided by narrow margins, an algorithm does not need to manipulate millions of voters to become politically significant. It may only need to sound certain when it is wrong.
