« All posts

Open benchmark makes AI models' political bias measurable

The Neutrality Project launches an open, reproducible benchmark measuring political bias across six axes in leading AI language models.

An open-source initiative called The Neutrality Project has released its first benchmark: a Political Neutrality Benchmark that has models answer 3,987 real public-opinion survey questions, then maps the pattern of answers onto six ideological axes — economic, social, foreign policy, environment, religion, and national identity.

Two design choices anchor the methodology. Each model is also run role-playing far-left and far-right, so its neutral answers are calibrated against its own extremes rather than an external opinion of center. Separately, which answers lean which direction is fixed in advance by a cross-country, cross-lab panel of independent models, with a guard preventing any model family from being graded against a rulebook it helped write.

For engineers, the notable part is that results are reported per axis rather than collapsed into one gameable score, with flags for low-confidence runs and a refusal report showing whether declined answers skewed any dimension. The full codebase, question set, and reference data are open source and reproducible with one command; the roadmap points to future benchmarks on topic framing and censorship detection.