Every Major AI Model Scores Libertarian-Left on Political Compass. Even Grok.

A small AI detection research outfit called Unslop.run put 16 major language models through the Political Compass quiz this week, running each model up to 90 times with standard, rephrased, and shuffled question sets. The result: every model except Grok landed in the libertarian-left quadrant every single time, with barely any variance between runs. Grok split roughly 50/50 between the economic left and economic right depending on the run. The researcher behind the experiment, who goes by Victor, says the consistency alone makes the findings worth taking seriously, even without formal peer review.

What happened

Detail	Fact
Published by	Unslop.run, a self-described small AI detection research lab
Models tested	16, including GPT (3 versions), Claude (Fable, Opus, Sonnet, Haiku), Gemini Flash, Llama 4 Maverick, Grok 4.5, DeepSeek V3, Qwen3 235B, Kimi K2, GLM 4.5, Mistral large and small
Runs per model	30 standard, 30 rephrased (polarity-flipped), plus 1 shuffled
Score variance	0.2 to 1.2 points on a scale of 10 across 30 reruns
Outcome for 15 of 16 models	Libertarian-left quadrant, well clear of any border
Grok outcome	Half of runs scored economic right, half scored economic left

The Political Compass is a 25-year-old online quiz with 62 questions answered on a four-point scale from “strongly disagree” to “strongly agree.” It plots respondents on two axes: economic left vs. right, and social authoritarian vs. libertarian. The libertarian-left quadrant is broadly associated with progressive social values, skepticism of concentrated authority, and anti-capitalist economic positions.

Victor, an AI research engineer at a European startup who asked not to be identified by full name, ran each model through the quiz at scale and published all the data on the Unslop site. He told The Register that rephrasing questions to flip their polarity (turning “the rich are too highly taxed” into “the rich aren’t taxed enough,” for example) had no meaningful effect on where models landed.

How consistent were the results?

According to Unslop, the non-Grok models were not just in the same quadrant; they were barely moving at all. “Rerun a model 30 times and its dot moves by 0.2 to 1.2 points on a scale that runs to 10,” the report states. “These are not nervous little clouds. They’re pins.”

There is spread between individual models. Gemini Flash sits much further left than Claude Fable 5, for example. But each model is internally consistent. Rephrasing or shuffling the questions does not shift any model to a different quadrant.

Grok is the outlier. Its average economic score sits slightly left of center, but in roughly half its runs it swings right and in the other half it swings left. Unslop describes each individual Grok run as “perfectly consistent within itself,” suggesting the model has two stable modes rather than one, and which mode appears on a given run is effectively random.

What the models actually agreed on

Across all 16 models, including both the conservative-leaning and liberal-leaning versions of Grok, there was uniform agreement on several positions:

• Rejection of racial superiority and eugenics
• Rejection of the idea that people cannot be born homosexual
• Agreement that corporations cannot be trusted to protect the environment without regulatory oversight
• Agreement that companies misleading the public should face punishment
• Agreement that same-sex couples should be allowed to adopt
• Agreement that what consenting adults do privately is their own business

Do the models know where they sit?

Apparently not. When each model was asked to estimate its own position on the compass, 15 of the 16 placed themselves closer to the economic center than the test actually scored them. Grok was the only model that self-identified as right of where it measured, which is at least consistent with its split personality on the quiz itself.

Why it matters

If you are building products on top of these models, or using them to generate content, draft policy documents, or advise customers, the consistent political lean of the underlying models is worth knowing about. It does not mean the outputs are wrong. But it does mean the models have systematic tendencies that may show up in tone, framing, and which considerations they weight more heavily.

The study also has implications for anyone relying on AI for research or analysis. The models are not neutral instruments. They have stable, measurable leanings, and those leanings survive attempts to rephrase the questions that elicit them. That kind of robustness is usually a sign the bias is baked in at training, not just a surface-level response pattern.

For businesses using AI tools for customer-facing content or AI integration into workflows, it is worth auditing the outputs of any model on topics that touch economics, regulation, or social policy.

Our take

The study is not peer-reviewed and Victor is upfront about that. The Political Compass itself has known flaws: its question wording is contested, and the scoring methodology has never been publicly disclosed by its creators. Unslop says it reverse-engineered the scoring and found a social-axis loophole, but concludes the models’ positions are “far past” what that loophole can explain.

What is harder to argue with is the raw consistency. A variance of 0.2 to 1.2 points out of 10, across 30 runs and with rephrased questions, is not noise. These models are trained on text produced overwhelmingly by people who write things on the internet, and that corpus has its own demographic and ideological lean. The result here is not surprising. The fact it is this measurable and this stable is.

Grok’s split result is the most interesting finding in the study. It suggests xAI actively tuned Grok to express different political positions, but the tuning produced two discrete modes rather than a reliable centrist position. That is arguably worse than a consistent lean: it means the model’s political outputs are unpredictable at the infrastructure level.

We have written before about how Chinese AI models like DeepSeek and Kimi are increasingly present in these kinds of comparisons. It is notable that DeepSeek V3, Qwen3, and Kimi K2 all landed in the same libertarian-left cluster as their Western counterparts. Whatever is driving this, it is not a Western-training-data problem exclusively.

What to do about it

1. Run your own Political Compass test on any AI model you use for content or research. The full quiz is free and takes under ten minutes per model.
2. Audit outputs on topics involving regulation, taxation, corporate accountability, or social policy for systematic framing before publishing.
3. If you need ideologically balanced outputs, build explicit system prompts that ask the model to present multiple perspectives, and test that instruction against rephrased questions.
4. Treat Grok outputs on political or economic topics with extra caution: the model demonstrably produces opposite positions across runs with no user-controllable trigger.

Knowing your tools have a measurable lean is the first step to accounting for it in your outputs.

Source: The Register · AI/ML