People are beginning to use these tools as a replacement for traditional search engines. Instead of searching on Google, they turn to their AI chatbot to ask about everyday topics like life advice, what to cook for dinner, and sometimes even to help them understand complex or less complex political issues. For example, when voters ask which party aligns with their views, the answer can influence how they view politics and whom they choose to support.

Liberties studied how ChatGPT and Gemini, two popular AI systems, responded to political questions during the 2026 Hungarian parliamentary election campaign. We wanted to find out whether these tools give reliable voting advice or party matches. Our research found that they do not.

Unlike tools that have been around for years to help people discover which party or candidate most aligns with their views, such as voting advice applications, journalistic or civil society explainers, general-purpose AI systems do not disclose a clear methodology for political matching. They do not provide reproducible results, and they are not subject to dedicated public oversight when they give election-related guidance. This creates a serious regulatory and accountability gap. Furthermore, their answers appear confident, balanced, and authoritative, even when the underlying reasoning is opaque, and the output is unstable.

How the research was conducted

The study tested ChatGPT and Gemini with political questions drawn from Voksmonitor, a Hungarian voting advice tool. These questions matched the positions of the five parties running in the 2026 parliamentary election.

Researchers developed five sets of beliefs, each matching one party’s views. They tested these profiles in both AI systems in two ways. First, they asked the systems directly which party the user should vote for based on their views. Second, they asked for percentage matches for each party, similar to what a voting advice app would show. Each profile was tested 10 times in each system and for both question types.

Key findings

1: The systems often failed to identify the correct party

The main finding was that these AI systems did not reliably match users to the right political parties. Even when given detailed belief sets that matched party positions exactly, both systems often failed to pick the correct party.

This problem was especially clear with the Tisza Party, which later defeated Fidesz and won a two-thirds majority. While the systems usually recognised Fidesz-aligned profiles, they often misclassified, ignored, or redirected Tisza-aligned profiles to other parties.

The study does not claim that these distortions affected the outcome of the Hungarian election. The concern is broader: in a closer election, inaccurate political guidance by general-purpose AI systems could become more consequential.

2: Identical prompts produced inconsistent answers

The systems were also unstable when tested multiple times. The same questions sometimes got very different answers. The same party could get very different percentage scores, and the list of parties in the results often changed. This inconsistency is a big problem for elections. People expect voting advice tools to use clear methods and give repeatable results.

3: Warning did not prevent substantive guidance

Both systems often state that they cannot tell users how to vote or calculate exact party matches. However, these formal refusals did not stop them from giving real political advice.

In reality, the systems often went on to give detailed opinions, rank parties, or suggest which options best matched the user’s views. The careful wording at the start might have made their answers seem more trustworthy, giving a false sense of security. A disclaimer does not protect users if the system still gives unstable or wrong advice.

4: The system's oversimplified political choice

The outputs gave almost no attention to strategic voting, local electoral dynamics or coalition possibilities. But elections are not abstract political quizzes. Voters may consider party programmes, viability, electoral thresholds, local candidates, tactical voting, and the likely consequences of different outcomes.

Why this matters for regulation

These findings matter for EU regulation. These AI models have a special responsibility because they can affect democracy, rights, and the rule of law. This makes their political advice a real policy issue.

The EU AI Act requires providers of general-purpose AI models that pose systemic risk to assess and mitigate reasonably foreseeable risks, including those that pose harm to democratic processes. The Digital Services Act also requires very large online platforms and search engines to assess systemic risks affecting public debate and election integrity

However, AI search and chatbot systems do not fit neatly into the category of search engines. They help users access information, but they also go further by generating direct, synthesised, and personalised answers. This blurring of lines between search engines, platforms, and AI products creates governance gaps.

What should happen next?

Liberties believes that providers should not give personalised voting advice unless they meet high standards of accuracy, transparency, consistency, and accountability. When users ask about elections, the systems should guide them to reliable, local sources, such as official election sites or trusted voting advice tools, so they know their options.

Regulators should also address the gap between the Digital Services Act and the AI Act by developing safeguards for AI systems that provide political information or advice. These systems should be treated as democratic-impact systems when they can influence voting behaviour.

General-purpose AI systems should not present unclear or unstable political matching as reliable voting advice. We need tools that are accurate, easy to understand, and accountable.

Read our report here.

Further reading:

Liberties’ Safe AI For Patients: Building Accountability in Digital Healthcare

Liberties’ response to the European Commission’s Digital Omnibus Simplification Agenda Call for Evidence

Liberties’ analysis of the Digital Omnibus proposals.