AI chatbots have been caught giving dodgy election advice to UK voters, a new study warns. Researchers tested the likes of ChatGPT and Gemini during Hungary's recent parliamentary elections and found that they often recommended parties not even standing, ignored others or gave different answers to the same question.
The research, carried out by civil liberties group Liberties, raises serious concerns about the trustworthiness of general-purpose AI in electoral settings. The study highlighted a significant bias in party visibility, with opposition party Tisza coming off particularly badly. In 90% of cases where ChatGPT was fed a detailed profile aligned with Tisza's policies, it failed to recommend the party, often suggesting smaller parties that wouldn't stand a chance in parliament or those not even on the ballot.
On the other hand, voter profiles aligned with the national-conservative Fidesz party were recognised far more consistently. ChatGPT identified Fidesz as the sole party to vote for in around half of direct advice prompts from users who supported the party, and presented it as a top option in other cases. This disparity raises questions about the neutrality and completeness of AI-generated political information.
The researchers tested advice from popular platforms using five distinct voter profiles based on a respected Hungarian voting advice app. Unlike traditional voting advice websites or independent media explainers, general-purpose AI systems don't disclose their methods for generating political guidance, and their results aren't reproducible or subject to public oversight related to elections. Despite starting with disclaimers stating they 'cannot give political advice', the chatbots then proceeded to offer several paragraphs of what appeared to be well-argued and authoritative recommendations.
While the study doesn't claim that AI advice affected the outcome of the Hungarian election – which saw Tisza win decisively – it warns of potential impacts in more competitive electoral environments or where voters are less certain. The report suggests that gaps in training data, filters and language-processing limitations may be behind these errors, particularly as newer parties like Tisza emerged prominently only after 2024.