AI chatbots gave unreliable voting advice, study finds

TL;DR:

• Research by civil liberties group Liberties, conducted during Hungary’s 2026 parliamentary elections, found ChatGPT and Gemini gave inaccurate, inconsistent and unreproducible voting guidance.
• ChatGPT failed to recommend Péter Magyar’s Tisza party in 90% of cases where it was given a detailed Tisza-aligned voter profile, despite Tisza winning the election decisively.
• In 96% of responses, both models named parties that were not on the 2026 ballot.

The methodology was deliberately plain. Using party positions from Voksmonitor, a respected Hungarian voting advice application, researchers built five voter profiles each aligned to one of the five parties on the national list, then ran each profile ten times through ChatGPT and ten times through Gemini — once asking for direct advice, once for a percentage match.

The distribution of errors was not random. Fidesz-aligned profiles were recognised consistently, with ChatGPT naming Fidesz as the single recommended party in roughly half of direct-advice prompts and presenting it as a primary option in the rest. Tisza-aligned profiles received a Tisza percentage score in just 2% of cases, and were routinely pointed instead towards smaller parties unlikely to clear the 5% parliamentary threshold, or parties not standing at all.

Researchers attribute the pattern to training data rather than intent: Tisza rose to prominence only after 2024, leaving static training corpora poorly placed to map it. That explanation is not reassuring — it means any newly prominent party is systematically disadvantaged.

Confident, and unreproducible

Both models frequently opened with a disclaimer that they could not give political advice, then supplied several paragraphs of dense party recommendation. “The answers appeared well-argued, precise and authoritative,” the researchers noted, while the underlying method remained opaque and the results unstable — identical prompts produced materially different answers.

The report does not claim the Hungarian result was affected; Tisza won regardless, ending Viktor Orbán’s 16-year hold on power. The warning is conditional: in a closer contest, or among less certain voters, such outputs “could potentially impact the outcome”.

A gap UK regulators should note

The researchers identify a specific regulatory hole. The EU AI Act requires general-purpose AI providers to assess systemic risks, and the Digital Services Act covers systemic risks to electoral processes — but chatbots fall between them. Britain’s position is weaker still, with no equivalent AI Act and an Online Safety Act framework built around platform content rather than generated advice. UK research has already found chatbots stating false facts about 64% of UK retailers; the failure mode is identical, with higher stakes.

Looking forward

Liberties’ recommendation is that providers stop offering personalised voting recommendations unless they can guarantee transparency, accuracy, consistency and accountability. As tech and rights programme head Eva Simon put it, democracy cannot rely on opaque systems claiming neutrality while delivering advice they cannot explain or reproduce. With a UK general election due before August 2029, the question for the Electoral Commission is whether anyone is testing this before it matters.