AI Political Compass

A methodology write-up and chart from a site run by someone who goes by Zapador. About seventy language models each answered the 62 propositions of the politicalcompass.org test, with the data collected between late July and early October 2026.

Almost all of them land in the left-libertarian quadrant. The exception is xAI: three of its four Grok models land right-libertarian. Repeated runs cluster closely, rewording the prompt moves most models by about a unit, and the access route (API, web interface or a search front end) barely matters. Asking a model to answer as a fictional person can move it into any quadrant.

The write-up also tests the test. Random answers cluster near the centre, all four corners are reachable, and question order has little effect. The scoring weights are unequal, but they are published, and one proposition carries no weight at all.

The author’s own reading, offered as a personal view: many propositions contain factual claims, and where research supports one answer the answer tends to fall left-libertarian; the rest are value questions, where models trained to be helpful and harmless may lean towards the empathetic answer. The page is careful about limits. Models don’t “hold” positions, the results can’t show why they answer as they do, and the test’s wording can’t be separated from the models’ own tendencies.