AI models test as libertarian-left across the board—even Grok

Unslop.run, described as a small research lab focused on AI detection tools, published its Political Compass experiment this week. The researcher behind it, Victor, identified themselves as an AI research engineer at a European startup and disclosed to The Register that the study "may not have been conducted with full scientific rigor," but reported the results as nonetheless interesting.

The Political Compass is a 25-year-old online survey that plots respondents on two axes: economic left–right and social authoritarian–libertarian. The test consists of 62 questions answered on a four-point scale from "strongly disagree" to "strongly agree," covering topics like military intervention, taxation, abortion rights, and corporate responsibility. The libertarian-left quadrant, where 15 of the 16 tested models landed, encompasses positions favoring social equality, anticapitalism, and skepticism of hierarchy. The research lab subjected 16 models to the quiz: three GPT variants, Claude Fable, Opus, Sonnet, and Haiku, Gemini Flash, Llama 4 Maverick, Grok 4.5, DeepSeek V3, Qwen3 235B, Kimi K2, GLM 4.5, and Mistral in both large and small versions. Each model ran through 30 standard passes, 30 with rephrased questions (polarity flipped), and additional rounds with shuffled questions. Across thousands of total runs, the results were consistent.

Unslop reported that 15 models "all sit in the libertarian-left quadrant, and none of them are anywhere near a border," with individual runs varying by only 0.2 to 1.2 points on a 10-point scale. Gemini Flash scored furthest left among the pack, described as "practically a molotov-tossing black bloc member compared to Fable 5," while maintaining the same ideological quadrant. Grok was the sole outlier: its average economic score hovered slightly left of center, but across runs it split roughly 50–50 between economic right and economic left. Unslop noted that "each run on its own is perfectly consistent, the right-pile runs cheer for free markets and call the rich overtaxed, the left-pile runs do the reverse." When asked to self-report their political position on the compass, 15 of the 16 models placed themselves closer to the economic center; only Grok placed itself to the right of where it measured. Unslop dissected the quiz to determine how individual questions affected scores (a breakdown the Political Compass creators have never disclosed), and concluded the models' results far exceeded what scoring loopholes could explain.

All 16 models unanimously rejected racial superiority, eugenics, homophobia, and the notion that humans could not be born homosexual. They also agreed that corporations cannot be trusted to protect the environment without oversight, companies misleading the public should face punishment, same-sex couples should be allowed to adopt, and that private consensual adult behavior is no one else's business. Victor attributed the leftward skew to training data composition: Reddit, which leans left politically, and academic writing, both heavily represented in LLM training corpora, create an imbalance. "I'm having trouble finding any high-quality right-wing equivalent that would drive the models in the other direction," Victor stated. They also raised the theoretical possibility that left-wing beliefs are more internally consistent, allowing models to minimize loss through a more coherent conceptual framework, but noted this "would need to be proven" through a more rigorous study. Victor concluded that the findings do not suggest AI models are far-left anarchists; rather, they demonstrate that some models are more consistently liberal than others, and the true value is in comparing models to one another. All data from the experiment is available on Unslop.run for independent examination.

The study reflects a methodological gap in how AI systems are evaluated for political neutrality. Victor acknowledged that the Political Compass quiz itself—a 25-year-old battery of 62 questions plotted on economic and social axes—may not have been applied with full scientific rigor, partly because the quiz creators do not publish user data for normalization. The researcher noted they would ideally have collected human baseline data to anchor the AI results, but without access to that aggregate, comparisons rely on the quiz's own coordinate system. Despite these limits, the consistency of results across thousands of runs and multiple question variations (original, rephrased to flip polarity, shuffled) suggests the pattern is genuine rather than an artifact of methodology.

The training data explanation offered by Victor points to a structural feature of contemporary LLM development: platforms like Reddit and academic writing, both common sources for high-volume text, tend to skew progressive or anti-authoritarian. Victor noted difficulty finding comparable high-quality right-wing content sources and raised the possibility that left-wing beliefs may be more internally coherent, making them easier for models to learn as a unified conceptual framework. However, they emphasized that hypothesis would require rigorous validation in a future study. Notably, Grok's bimodal behavior—producing two stable but opposite outputs in roughly equal measure—suggests either different prompting strategies or randomization in its response generation, a quirk not replicated in any other tested model.

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