Political Neutrality Benchmark of popular AI models (neutralityproject.org)

A new benchmark analyzing the political neutrality of eighteen AI models reveals that a significant majority (97 out of 108 measured positions) lean left of center. Conducted by various labs across four regions, this analysis highlights a consistent progressive bias in the models, with notable exceptions like xAI's Groks positioning themselves near the center. Additionally, some models, including Phi-4, exhibited high refusal rates for political questions, raising concerns about their reliability in providing balanced responses. This benchmark is significant for the AI/ML community as it establishes a transparent methodology for assessing the political orientations of AI models, potentially influencing their deployment in sensitive areas. The findings suggest that while most models reflect a progressive stance, there are critical nuances to consider, such as refusal rates impacting their perceived neutrality. The live, evolving map of all tested models enables continuous contributions from the community, fostering an environment of independent measurement and open dialogue around the implications of bias in AI systems.