An international research team including the University of Bayreuth has, for the first time, analyzed the “inner workings” of AI language models when anticipating political voting decisions. To do so, the researchers examined 6-national elections and AI-based election forecasts and evolved a new approach for making more exact predictions. They introduced their findings at the International Conference on Machine Learning (ICML 2026) in Seoul, South Korea.
AI-supported opinion research is already being utilized in academia, market research and political consulting. As such forecasts become increasingly incorporated into decision-making techniques, it is becoming ever more vital to understand the premise on which these predictions are generated.
Analyzing the inner representations of AI models supports shed light on the AI “black box” and disclose what information is really stored within the models and the way it is used. This makes AI forecasts more obvious and accurate, allows the identification of forecasting errors and supports uncover potential biases.
Large language models (LLMs) are more and more being used to evaluate attitudes, customer behavior and political preferences, as well as to forecast future developments. Moreover, previous studies of preference prediction have targeted primarily on the very last responses generated through LLMs instead on the process by which which those responses are produced.
A research team from Ludwig Maximilian University of Munich and the University of Bayreuth set out to address with this issue. The study is published at the arXiv preprint server.
“Previous research approaches that focused totally on evaluating AI-generated answers are instead like seeking only at the result showed via a calculator without understanding how the calculation was performed. In our study, we were efficiently capable of look over the AI’s shoulder as it was thinking,” stated Simeon Allmendinger, a doctoral researcher in the University of Bayreuth Information Systems and Human-Centered Artificial Intelligence research group.
In the study, the researchers evaluated more than 24 million integration of factors—which includes the specific language model used, individual demographic characteristics, party constellations and prompts (so-referred to as configurations)—across 7-language models and 6 national elections.
Within these different configurations, they investigated which internal regions of the LLMs were activated, how political parties were related with specific characteristics within the models and how information such as of age and educational background was processed.
“We determined that the ‘inner workings’ of language models frequently contain extra information that isn’t fully showed within the final answer. Making this information visible using of our new technique can offer extra insights and enhance forecast accuracy,” stated Professor Dr. Niklas Kühl, chair of Information Systems and Human-Centered Artificial Intelligence at the University of Bayreuth.
For example, if a language model is requested to anticipate which party a person could vote for and its answer is “Party X,” the model may also internally include indications that the person has an equally robust association with Party Y.
“Our outcomes reflects that the models learn more about underlying relationships than they in the end disclose of their final output,” Kühl describes. “It is vital to highlight that this technique is planned as a complementary tool and not as a alternative for traditional surveys. Mainly when it comes to underrepresented groups, real-world surveys stay indispensable,” Allmendinger stresses.
The study was carried out in collaboration with the Munich Center for Machine Learning (MCML), the Fraunhofer Institute for Applied Information Technology FIT and the University of Maryland.
