Political Compass Puts 15 of 16 LLMs in One Quadrant

An independent researcher ran the Political Compass test on 16 large language models and found 15 of them landing in the same spot, per unslop.run.

Fifteen of sixteen models return the same political lean, so asking a different chatbot for a second opinion doesn't actually get you one.

Fifteen models, from Google, Anthropic, OpenAI, Meta, Mistral, Qwen and Kimi, clustered at roughly (-6.3, -6.4) on the compass, solidly in the libertarian-left quadrant. All sixteen models, Grok included, agreed on 42 of 62 propositions every run.

Grok is the exception. Instead of settling on one position, it split 50/50 between two personas: one left-leaning at -5.9 on the economic axis, one right-leaning at +3.3. Same model, two different politics.

The researcher ran 30 standard passes per model, 30 with the questions reverse-phrased, and 10 with the answer order shuffled. That's about 69,440 answers, with the scoring system reverse-engineered to control for ordering effects and sycophancy.

That kind of agreement across labs that don't share training pipelines looks less like consensus and more like one shared blind spot everyone inherited. Watch whether any lab deliberately widens that range, or whether Grok's split turns out to be a design choice rather than a fluke.

The reverse-phrasing and replication controls, out July 27, are what separate this from a one-shot bias post.