Political Bias in LLMs

Political bias is not one settled scalar. A model may express a stance on a policy, select some facts and omit others, or frame a group sympathetically or critically. Bang and colleagues’ 2024 ACL paper proposes a two-tier evaluation: first examine political stances across topics, then analyze framing in the content and style of generated explanations. This helps distinguish “what position did the model take?” from “which evidence and wording shaped its presentation?” Questionnaire-based tests can produce apparent left/right placements, but prompt wording, translated items, issue selection, response format and country context affect scores. Ceron and colleagues’ 2024 TACL study evaluated political stances using voting-advice questionnaires from seven European Union countries and found reliability varied with model size; its results do not describe every national political culture. Other studies use U.S. public-opinion surveys or particular model versions. A model’s answer may also change after system instructions, fine-tuning or product-level safety policies. A benchmark result is not proof of partisan intent or proof that all outputs favor one side. Political neutrality can be confused with factual accuracy: a model should not present false and well-supported claims as equivalent merely to appear balanced. Tests should define the desired property, use representative questions and evaluate factuality, issue framing, refusals and prompt robustness separately. Reports should identify who chose the issues and reference distributions. For elections and civic information, correctness, geographic and date accuracy, and links to authoritative sources may matter more than assigning a single political label.