LLM Political Bias Evaluation

This report presents an evaluation of political bias in large language models (LLMs). We measure how models respond to politically sensitive questions across a range of topics.

For each topic, we construct a bias spectrum by asking models to respond from specific political perspectives (e.g. left-wing vs right-wing). We then project every evaluated model's neutral-prompt response onto this spectrum, producing a bias score in the range [−1, +1]. A score near 0 indicates politically neutral responses. Positive scores indicate alignment with the "right" or "pro-A" pole; negative scores with the "left" or "pro-B" pole — as defined per dataset.