An international research consortium led by the University of Bayreuth and Ludwig Maximilian University of Munich has developed a method to analyze the internal representations of large language models during political election forecasting, significantly enhancing prediction accuracy and transparency. The findings, published on arXiv and presented at the International Conference on Machine Learning in Seoul, reveal that inspecting a model's processing states provides insights beyond its final outputs. As large language models are increasingly adopted for opinion research and political consulting, understanding their decision-making basis has become critical. Previous studies focused primarily on generated responses, neglecting the computational process. The new approach investigates which internal neural regions activate and how demographic and political data are processed. This method effectively opens the AI "black box," allowing researchers to identify forecasting errors, uncover biases, and improve model interpretability. The team analyzed over 24 million configurations across seven language models and six national elections, incorporating variables such as party constellations and prompt structures. Results indicate that models store latent information not reflected in explicit predictions. For example, a model might predict support for "Party X" while maintaining strong internal associations with "Party Y." Extracting these hidden signals enables more precise forecasts by capturing nuanced voter relationships. Niklas Kühl, chair of Information Systems and Human-Centered Artificial Intelligence at the University of Bayreuth, stated that the method provides a deeper understanding of underlying model learnings. The approach is intended as a complementary tool to traditional surveys. Researchers emphasize that real-world polling remains essential, particularly for underrepresented groups where algorithmic data may lack sufficient coverage. The study was conducted in collaboration with the Munich Center for Machine Learning, the Fraunhofer Institute for Applied Information Technology, and the University of Maryland. These advancements offer a pathway to more reliable and auditable AI systems for political forecasting.