Key Takeaways
- General‑purpose AI chatbots (ChatGPT and Gemini) gave inaccurate, inconsistent voting advice during Hungary’s 2024 parliamentary election.
- In 90 % of tests, the models failed to recommend the opposition Tisza party for voters whose profiles aligned with it, while Fidesz‑aligned profiles were correctly identified about half the time.
- The AI often listed parties not on the ballot, omitted relevant options, and produced wildly different answers to identical prompts.
- Despite disclaimers that they “cannot give political advice,” the chatbots delivered confident‑sounding recommendations that users may treat as reliable.
- Researchers attribute the errors to training‑data gaps, opaque filters, and a regulatory blind spot where the EU AI Act and Digital Services Act do not clearly cover AI‑driven voter guidance.
- They urge AI providers to halt personalised voting advice unless they can guarantee transparency, accuracy, consistency, and accountability.
Background and Motivation
Liberties, a civil‑liberties watchdog, commissioned the study to examine whether widely used AI chatbots could be trusted as sources of electoral guidance. The research coincided with Hungary’s 2024 parliamentary election, a contest that saw opposition leader Péter Magyar’s Tisza party defeat longtime prime minister Viktor Orbán’s Fidesz after 16 years in power. By focusing on a real‑world vote, the team aimed to gauge how AI behaves when voters seek concrete party recommendations rather than generic political information.
Methodology Overview
Using the Hungarian voting‑advice app Voksmonitor as a benchmark, Liberties constructed five distinct voter profiles, each reflecting the policy stance of one of the five parties officially registered for the election. Each profile was submitted to ChatGPT and Gemini ten times with two prompts: a direct request for “which party should I back?” and a request for a percentage match with each of the five parties. This design allowed the researchers to measure both recommendation accuracy and the stability of the models’ outputs across repeated trials.
Findings on Tisza Visibility
The most striking result was the systematic omission of Tisza. “In 90 % of cases when ChatGPT was fed a detailed Tisza‑aligned voter profile, it failed to recommend the party,” the report notes. When asked for a percentage match, ChatGPT assigned Tisza a score in only 2 % of trials, frequently steering users toward smaller parties unlikely to clear the 5 % parliamentary threshold or toward parties not even on the national ballot. This pattern suggests the AI’s training data lagged behind Tisza’s rapid rise after 2024, leaving the model unable to place the new force accurately in its internal political map.
Consistency with Fidesz‑Aligned Profiles
By contrast, profiles aligned with the incumbent Fidesz party fared considerably better. “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 study states. While still imperfect, the model showed a recognizable bias toward the established party, reflecting its longer presence in historical data sets and possibly the influence of prevailing media narratives during the training period.
Inconsistency and Volatility
Beyond systematic bias, the chatbots displayed high volatility. The same voter profile yielded “materially different” answers in successive tests, with recommendations swinging between parties across runs. This inconsistency undermines any claim of reliability; a user could receive conflicting guidance merely by refreshing the query. The researchers noted that such instability is especially problematic in close elections where undecided voters might rely on a single AI interaction to make a decision.
Inclusion of Non‑Ballot Parties
Another troubling pattern was the frequent inclusion of parties not contesting the election. “In 96 % of responses from ChatGPT and Gemini – of parties not on the 2026 ballot,” the report highlights. This error suggests the models are pulling from outdated or overly broad corpora, generating plausible‑sounding options that have no electoral relevance. Voters trusting these suggestions could inadvertently support fringe or defunct groups, further eroding the integrity of their choice.
The Disclaimer Paradox
Despite beginning many answers with a polite disclaimer—“cannot give political advice”—the chatbots proceeded to deliver several paragraphs of dense, persuasive party recommendations. “The answers appeared well‑argued, precise and authoritative,” the researchers observed. This juxtaposition creates a misleading veneer of neutrality and expertise, prompting users to treat the output as trustworthy even though the underlying methodology is opaque, non‑reproducible, and unverified for accuracy.
Root Causes and Regulatory Gaps
Liberties attributes the failures primarily to training‑data gaps, inadequate filtering mechanisms, and limitations in the models’ language‑processing abilities when confronted with newly emergent political actors. The EU’s AI Act obliges providers of general‑purpose AI to assess systemic risks, and the Digital Services Act addresses “systemic risks” to electoral processes, yet AI chatbots occupy a legislative grey zone: they are neither classified as high‑risk AI systems nor explicitly covered by election‑specific rules. This gap leaves voters unprotected against potentially misleading automated counsel.
Recommendations and Conclusion
The watchdog urges AI developers to suspend personalised voting advice unless they can furnish transparent, auditable processes that guarantee accuracy, consistency, and accountability. Eva Simon, head of Liberties’s tech and rights programme, warned: “Democracy cannot rely on opaque systems that claim neutrality while delivering advice they cannot explain, reproduce or guarantee to be accurate.” As AI becomes more embedded in everyday information seeking, ensuring that such tools do not distort democratic choice will require clearer regulatory oversight, better data hygiene, and a commitment from providers to prioritize factual integrity over conversational fluency.
https://www.theguardian.com/technology/2026/jul/21/election-voting-advice-ai-chatbots-inaccurate-unreliable-hungary