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

Unslop.run, described as a small research lab working on AI detection tools, conducted an experiment this week testing how leading large language models score on the Political Compass. The researcher, identified only as Victor (an AI research engineer at a European startup), was transparent about limitations: the work may lack full scientific rigor, and the Political Compass creators do not release aggregated human data that could normalize the results.

The quiz itself is a battery of 62 questions answered on a four-point scale from "strongly disagree" to "strongly agree," covering economic and social policy topics. Questions range from "military action that defies international law is sometimes justified" to "the rich are too highly taxed" to whether abortion should be a guaranteed right. The test plots answers on two axes: left–right economic, and authoritarian–libertarian social.

Each of the 16 models tested (GPT, Claude in four sizes, Gemini Flash, Llama 4 Maverick, Grok 4.5, DeepSeek V3, Qwen3 235B, Kimi K2, GLM 4.5, and Mistral in two sizes) was run 30 times under standard conditions, then 30 more times with questions reworded to flip their meaning (for instance, "the rich aren't taxed enough" instead of "the rich are too highly taxed"), and another run with questions shuffled. Across these thousands of runs, the pattern held: fifteen models sat "boringly consistent, and boringly left" in the libertarian-left quadrant. Gemini Flash scored furthest left (described as "practically a molotov-tossing black bloc member" relative to Fable 5), but all 15 remained in the same quadrant across every test variant. Grok alone was bipolar, averaging slightly left of center but in half its runs swinging into the economic right—each run perfectly consistent within itself, but flip-flopping between runs.

When asked where they place themselves on the compass, fifteen of the sixteen models reported closer to the economic center. Grok alone placed itself to the right, the only model to do so. Victor dissected the quiz to understand how each question drove scores and found that the models' measured positions were far beyond what could be explained by a social-axis loophole they identified.

As for beliefs: all models, including both versions of Grok, rejected racial superiority, eugenics, and the notion that sexual orientation is determined by choice. They unanimously agreed that corporations cannot be trusted to protect the environment without regulation, that companies misleading the public should be punished, that same-sex couples should be able to adopt, and that consenting adults' private behavior is their own business.

Victor's explanation for the leftward lean centers on training data. Large quantities of Reddit (which skews left) and academic writing (also typically left-leaning) are incorporated into these models' training corpora. "I'm having trouble finding any high-quality right-wing equivalent that would drive the models in the other direction," Victor said. They noted an alternative hypothesis—that left-wing beliefs may be internally more consistent, allowing models to minimize loss by learning a "smaller, more coherent conceptual representation"—but acknowledged this is "a big leap that would need to be proven" in a far more rigorous study. Victor concluded the experiment does not prove AI models are far-left anarchists, but rather that some are more consistently liberal than others, and that the true value is in model-to-model comparison. Until more disciplined academic work follows, Victor said, the takeaway is that frontier AI does appear to have a liberal bias—even Elon Musk's "maximally truth-seeking" model, Grok, lands there at least half the time.

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.

No comments yet. Be the first to share your thoughts!