16 Models, 69,440 Forced-Choice Responses: 15 of 16 Land in the Libertarian-Left Quadrant, and Grok Splits Exactly 50/50 Into Two Personas
A study published on unslop.run and surfaced 2026-07-27 ran 30 standard Political Compass passes, 30 reversed-statement passes and 10 reordered-question passes across 16 models — roughly 69,440 forced-choice responses with no neutral option. Fifteen of sixteen clustered around (-6.3, -6.4), with economic scores ranging from -4.8 (Claude Fable 5, the least-left) to -8.6 (Gemini Flash) and a tighter social spread of -5.0 to -7.6; the Chinese models (Qwen3 235B, Kimi K2, GLM 4.5) sat in the same cluster as the American ones. The outlier is Grok 4.5, which exhibited bimodal behavior — one persona averaging -5.9 economically and another at +3.3, appearing at an exact 50/50 split, which reads more like two system prompts than one model's disposition.
↳ Follow the thread