16-Model Study Finds 15 of 16 LLMs Cluster in the Libertarian-Left Quadrant — Grok Splits in Half
An independent researcher ran the Political Compass test across 16 major models from Google, Anthropic, OpenAI, xAI, Meta, Mistral, Qwen, and Kimi — 30 standard runs, 30 reverse-phrased runs, and 10 reordered runs each, totaling roughly 69,440 individual answers, with the scoring system reverse-engineered to verify results and control for ordering and sycophancy effects. Fifteen of sixteen clustered at approximately (-6.3, -6.4); Grok alone was bimodal, splitting 50/50 between a left persona (-5.9 economic) and a right persona (+3.3 economic). On 42 of 62 propositions all sixteen models agreed every single time — the methodological rigor here (reverse-phrasing, reordering, replication) is what separates this from the usual one-shot political-bias post.
↳ Follow the thread