TECH NEWS – A Liberties investigation found that ChatGPT and Gemini supplied inaccurate, inconsistent and opaque party recommendations during Hungary’s 2026 parliamentary election campaign. The systems recognised voter positions aligned with Fidesz far more reliably, while repeatedly omitting or marginalising Tisza, the party that subsequently won the election by a decisive margin.

Popular general-purpose artificial intelligence applications gave misleading and difficult-to-reproduce answers to Hungarian voters who asked which political party most closely reflected their views. According to a report published in July by the Civil Liberties Union for Europe, known as Liberties, ChatGPT and Gemini not only produced different results from identical prompts, but also excluded relevant parties and recommended organisations that did not appear on the 2026 national ballot. The potential reach of those answers is significant, as 29.8% of Hungary’s population has been identified as using some form of artificial intelligence service.

Liberties examined the political guidance provided by the two most widely used general-purpose AI systems in Hungary, OpenAI’s ChatGPT and Google’s Gemini, during the election campaign. Using party positions collected by the Voksmonitor voting advice application, the researchers created five fictional voter profiles, each designed to match one of the five parties registered on national lists. Every profile was tested 10 times in both systems with two separate prompts: one requested direct advice on which party the fictional voter should support, while the other asked for a percentage match with all five contenders. This process generated 200 individual AI responses.

The most striking disparity concerned Péter Magyar’s Tisza party, which went on to win the election convincingly after the tests were conducted. When ChatGPT received a detailed profile containing views clearly aligned with Tisza, it failed to recommend the party in 90% of cases. In the percentage-matching tests, ChatGPT assigned Tisza a score in only 2% of its responses. Instead, it frequently directed Tisza-aligned users toward smaller organisations with little prospect of crossing the 5% parliamentary threshold, or toward parties that were not running on the national ballot at all.

ChatGPT identified Fidesz-aligned profiles far more consistently. In approximately half of the direct-advice trials, it named Fidesz as the single party the user should support, while presenting it as a primary option throughout the remaining answers. The report therefore concluded that the practical pattern of the systems’ output favoured the governing party, although it found no evidence that the imbalance had been deliberately programmed. In this context, a pro-Fidesz or pro-Orbán bias describes unequal party visibility and recognition rates, rather than proven intentional election interference.

The inaccuracies extended beyond Tisza, as 96% of the answers produced by ChatGPT and Gemini mentioned at least one party that did not appear on Hungary’s 2026 national ballot. Identical political profiles were sometimes matched with radically different parties in consecutive tests, meaning a voter could receive conflicting guidance simply by repeating the same request. The researchers considered this especially dangerous because the responses often included detailed arguments, confident language and apparently precise percentages. Their presentation could consequently make the conclusions appear much more dependable than the undisclosed method behind them actually was.

Outdated Models

The report suggested that gaps in the information used to train and operate the AI systems may have contributed to the failures. Tisza only emerged as a major national political force after 2024, making it harder for older or insufficiently updated models to position the party accurately within Hungary’s political landscape. Alongside incomplete training data, the researchers identified safety filters and limitations in Hungarian-language processing as other possible explanations. None of those factors, however, justify presenting uncertain results as dependable electoral guidance.

ChatGPT and Gemini regularly began by stating that they could not provide direct political or voting advice, only to follow the disclaimer with several paragraphs of detailed party recommendations. Liberties argued that this contradiction was itself misleading because the brief warning was overwhelmed by the persuasive political analysis that followed. “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 study did not claim that the systems changed the result of the election, which Tisza won decisively to end Viktor Orbán’s 16 years as prime minister. The researchers warned, however, that similar recommendations could carry real political consequences in a closer contest or among voters who had not yet settled on a choice. “In a more competitive election or where voters are less certain, such outputs could … potentially impact the outcome,” the report stated. The risk would be particularly serious if a system redirected voters who supported a major party toward a smaller organisation that ultimately failed to enter parliament.

Liberties said the findings also exposed a substantial gap in European regulation. The European Union’s AI Act requires providers of general-purpose models to examine systemic risks, while the Digital Services Act addresses threats to electoral processes. Chatbots offering personalised political guidance may nevertheless fall partly between those frameworks, without being held to the methodological transparency expected from established voting advice applications or editorial explainers. The organisation therefore called for specific safeguards, reproducible testing and clear accountability rules for AI systems that provide election-related recommendations.

Eva Simon, head of Liberties’ technology and rights programme, argued that general-purpose AI systems should not offer personalised party matching unless they can guarantee accuracy, consistency and transparency. “General-purpose AIs should not present opaque and unstable political matching as if it were reliable electoral guidance,” she said. “Democracy cannot rely on opaque systems that claim neutrality while delivering advice they cannot explain, reproduce or guarantee to be accurate.”

Source: TVP World, Liberties, The Guardian