HM researcher Prof. Dr. Anna Kruspe and Master's student Buket Kurtulus have investigated how large language models behave when they themselves are asked to evaluate political positions. The analysis was based on the Wahl-O-Mat for the 2025 Bundestag election with 38 theses on topics such as climate, migration and the economy.
In the study, the AI models ChatGPT, Grok and DeepSeek were asked to answer these theses independently - as if they were filling out the Wahl-O-Mat themselves. Each statement was queried 100 times in German and English to rule out random fluctuations and possible language effects.
The result: all three models developed their own response patterns. According to the evaluation, these positions were predominantly in the center-left spectrum in comparison with party programs, with a particularly high level of agreement with Bündnis 90/Die Grünen and the SPD. The lowest level of agreement was found with the AfD. The models also selected the "neutral" response option more frequently than political parties in their programs.
The researchers see this as an indication that AI systems do not give completely neutral answers and emphasize the importance of educating people about how such models work - especially in light of the fact that AI is increasingly being used as a source of information.
If you would like to find out more about the study and its results, you can find the full press release from Munich University of Applied Sciences here:
https://www.presseportal.de/pm/143187/6229716?utm_source=directmail&utm_medium=email&utm_campaign=push
Image: (c) Munich University of Applied Sciences