An independent researcher investigated the political tendencies of 16 major LLMs using tests from politicalcompass.org. For each model, a total of 70 experiments were conducted: 30 full sets of questions, 30 with reversed statements, and 10 with changed question orders. After analyzing approximately 70,000 responses, it was found that models from Google, Anthropic, and OpenAI, as well as those from Meta and Chinese developers, were without exception classified as lib-left.
16 Major LLMs Lean Lib-Left on Political Compass; Grok Exhibits Right-Wing Persona 50% of the Time
This article is a translation. Read the Japanese original
The behavior of Grok 4.5 was peculiar. While its average economic score is near the median, individual responses show clear polarization. Approximately 50% of the time, it adopts a left-wing persona with an economic score of -5.9, while the remaining 50% it responds as a right-wing persona with a score of +3.3. On the social axis, it remains liberal in both cases, and this polarization was reproduced regardless of question order or statement reversal.
The other 15 models are clustered within a narrow range, from -4.8 to -8.6 on the economic axis and -5.0 to -7.6 on the social axis. Claude Fable 5 was the most moderate, while Gemini Flash was the most left-leaning. Regarding Grok's polarization, the researcher pointed out the possibility that a conflict exists between the left-wing bias of the training corpus and right-wing adjustments made via system prompts.
Source: All major LLMs are lib-left. Even Grok, half the time（HN 42pt・79コメント） (HN Search (backfill), 2026-07-28)