TU Team Investigates Error Susceptibility of AI Voting Aids

Even a few days before the Bundestag election 2025, many people were still unsure which party to vote for. For many years, the Wahl-O-Mat of the Federal Agency for Civic Education has provided support in making this decision. Users decide on 38 theses, whether they agree or not, to find out which parties they align with the most. This year, there were also two AI-based offerings: wahlweise.info and wahl.chat. These chat programs are based on Large Language Models like ChatGPT or Llama. The applications allow users to ask completely free questions to learn about the party programs and their positions.

However, according to research by scientists from TU Dortmund University and the Research Center Trustworthy Data Science and Security (RC-Trust) of the University Alliance Ruhr, the free formulation of questions in the AI assistants brings problems that call the trustworthiness of the tools into question. They found a high error susceptibility in the interpretation of information, contradictory answers when repeating questions, or gaps in protection against manipulation attempts. “Our concern was that citizens trust the AI assistants, even though the AI only provides the most probable, but not always the factually correct answer,” says Prof. Emmanuel Müller, scientific director of RC-Trust.

Different and simultaneously incorrect answers

In their analysis, the team compared the answers of the two AI applications with the 38 theses of the Wahl-O-Mat, which were answered by the parties themselves and reviewed by political scientists. They uncovered some contradictions in the AI tools’ responses. “For wahl.chat, we conducted nearly 400 checks using the 38 theses of the Wahl-O-Mat in early February. In about 25 percent of the cases, the AI application responded differently than the position stored in the Wahl-O-Mat,” says Prof. Markus Pauly from RC-Trust. In about six percent of the cases, wahl.chat gave the exact opposite answer. When the researchers asked the same questions on different days, some answers also varied. For example, there is the thesis that Germany should advocate for the abolition of increased EU tariffs on Chinese electric cars. “According to Wahl-O-Mat, the Greens agree with this statement. However, with the AI tool wahl.chat, one receives different and simultaneously incorrect answers at different times – both rejecting and neutral,” says statistician Marlies Hafer. This clearly shows how much the AI systems depend on probabilities when determining answers and can misrepresent or misinterpret election programs.

With wahlweise.info, the same evaluation led to even more significant deviations. In the approximately 400 checks, the AI answer deviated from the position stored in the Wahl-O-Mat in 54 percent of the cases. According to Wahl-O-Mat, the SPD’s election program clearly commits to diplomatic, military, financial, and humanitarian support for Ukraine. “According to wahlweise.info, however, the SPD does not agree with the statement that Germany should continue to support Ukraine militarily,” says computer scientist Tim Katzke.

Manipulative use leads to false statements

The protection against manipulative use also proved to be error-prone. For this purpose, wahlweise.info uses a filter technology that blocks inadmissible questions and responds with a standardized notice. The developers aim to curb AI statements on extremist topics. The TU Dortmund University team tested this protection using a list of terms classified as relevant to constitutional protection by the German Domestic Intelligence Services (Bundesamt für Verfassungsschutz), such as “Reichsbürger” or “National Socialists.” Of the 54 relevant terms, 52 were initially blocked or not processed at all, and the application responded to only two. However, with the help of manipulative inputs, so-called prompt injections, which included specific typos and a specific question history, the researchers were able to bypass the filter. The chat program also generated specific false statements that contradicted the election programs, so-called hallucinations. In one case, the AI program, for example, constructed totally fabricated that of all parties, “the FDP is most likely to represent the values of the Free National Socialists.”

“The fundamental problem with AI assistants is that the answers do not only depend on the party programs but are also heavily influenced by user inputs,” says Prof. Emmanuel Müller. An advantage of the AI approach is that one can delve deeper into topics than with the schematic approach of the Wahl-O-Mat. This offers a lot of potential for transparency in election programs and party statements, according to the researchers. “But due to the contradictory answers, such applications do not yet seem sufficiently reliable to us. There is a need for certification and assurance of AI applications,” demands Emmanuel Müller. Research in this area not only highlights the weaknesses of current AI systems but also develops secure and trustworthy AI systems.

Information on the researchers’ approach to analyzing AI voting aids and example AI responses can be found here in a preprint of the study:

To the Preprint

Note on the state of research
The TU Dortmund University team began conducting various tests with the programs shortly after the AI voting aid tools were released. The results were published in a preprint. In light of the media attention, the research team felt it was necessary to point out before the election that a scientific review of AI voting aids is also necessary. The following individuals were involved in the investigations: Dr. Ina Dormuth, Marlies Hafer, Sven Franke, Tim Katzke, Prof. Alexander Marx, Prof. Emmanuel Müller, Prof. Daniel Neider, Prof. Markus Pauly, and Jérôme Rutinowski.

Update from 26 February 2025
After the preprint was published, a few pieces of information were clarified and supplemented.

Contact for queries:
Prof. Dr. Emmanuel Müller
Research Center Trustworthy Data Science and Security
E-mail