AI models test as libertarian-left, even Grok flips sides

The Political Compass is a 25-year-old online quiz that plots respondents on two axes: economic (left to right) and social (authoritarian to libertarian). The libertarian-left quadrant represents views favoring social equality and anticapitalism alongside a disdain for hierarchies—territory typically associated with anarchists, libertarian socialists, and progressive movements. This is where 15 of the 16 tested models landed consistently.

The consistency itself is striking: when Unslop reran models multiple times, they reported that dots moved by only 0.2 to 1.2 points on a 10-point scale—"not nervous little clouds. They're pins," as the researchers put it. This stability across rewording, polarity flips, and question shuffles suggests the lean is not a quirk of phrasing but something deeper in how the models process political concepts.

Grok's behavior stands apart. It split cleanly into two populations: one set of runs placed it left, the other right, with no middle ground. Victor observed that in right-leaning runs, Grok "cheer[s] for free markets and call[s] the rich overtaxed," while in left-leaning runs it does the reverse. Yet even Grok, when asked directly where it places itself, answered to the right of center—the only model to do so. All 15 other models self-reported closer to economic center, even though the compass test showed them consistently left. This gap between self-perception and measured output is a core finding: the models appear unaware of their own political lean.