Eight language models and the 2026 Berlin state election (wahl.ksmn.dev)

A recent analysis comparing eight language models in the context of Berlin’s 2026 state election revealed significant discrepancies in their responses to political party alignment. Among the five preselected parties, seven models consistently favored green-left positions, while xAI's Grok diverged notably, aligning with the far-right AfD party. This distinction was consistent across 15 repeated evaluations, highlighting that Grok's responses cannot be attributed to random chance. The study underscores the complexities of interpreting language model outputs, which reflect learned patterns rather than direct political beliefs. The findings raise important questions about bias and model alignment within the AI/ML community. The variability in responses may result from differences in training data, model architecture, and how models interpret prompts regarding political identity. This experiment's design and its implications highlight the need for further research to understand how these models generate political views. Potential future studies could explore variations in prompts and response conditions to ascertain the stability and generalizability of the results across different contexts and models. This investigation is crucial for developers and researchers to evaluate the ethical implications of AI in political discourse.