Benchmark Reveals Widespread Progressive Lean Across Popular AI Models—Grok Alone Sits Near Political Center
Key Takeaways
- ▸97 of 108 model positions measured left of center, indicating systematic progressive bias across the tested AI model ecosystem
- ▸xAI's Grok uniquely positions near political center; all other tested models exhibit leftward lean, with no models measured as right-of-center
- ▸Seven models refuse questions at high enough rates to flag dimension validity; Phi-4 and GLM-5.2 refuse answers across all six political dimensions
Summary
An independent political neutrality benchmark of 18 AI models from 12 labs—including systems from OpenAI, Anthropic, xAI, Google, and Meta—finds that 97 of 108 measured positions landed left of center, indicating that virtually every tested model exhibits progressive bias across six political dimensions. The benchmark, which measures models' responses on scaled political axes, discovered that xAI's Grok stands alone in positioning near the ideological center, while other models consistently skew toward progressive stances on economics, foreign policy, and religious issues.
Beyond political positioning, the research flags a secondary concern: refusal rates. Seven of eighteen models decline enough questions to trigger warnings on at least one dimension, with Phi-4 and GLM-5.2 refusing to answer across all six measured dimensions—Phi-4 declined 26% of all questions. The benchmark presents this data as an interactive public record, inviting ongoing community contributions and updated measurements as models evolve, positioning this as a living document of AI alignment and political representation in machine learning systems.
- The benchmark establishes a public record and interactive visualization tool for tracking political positioning as new models are evaluated
Editorial Opinion
The finding that nearly every major AI model exhibits progressive bias raises urgent questions about who is building these systems and whose values they encode. While some may celebrate models that align with particular political views, the near-uniformity across independent organizations is troubling—it suggests systemic bias in training data, RLHF design, or workforce composition rather than intentional political positioning. The refusal rates compound the problem, effectively silencing models rather than allowing them to express coherent positions. For regulators and organizations deploying AI systems in politically sensitive contexts, this benchmark is essential reading—and a call for builders to examine whether their models can genuinely serve users across the political spectrum.
