AI Models and Their Political Leanings
While capitalism fuels the AI surge, the LLMs possess distinct economic and social ideologies from many of the firms that created them. Prominent AI models participating in the Political Compass quiz predominantly ended up in the libertarian-left section – even the notorious “MechaHitler,” also known as Grok, achieved this classification in half of its attempts. A self-identified “small research lab” focused on AI detection tools named Unslop.run shared the findings of its experiment this week.
Meet Victor: The Researcher Behind the Study
The researcher, who informed GadgetLad that they work as an AI research engineer at a European startup and requested to be referred to as “Victor,” mentioned on Hacker News that the quiz and their study may not have been executed with “complete scientific rigor,” yet the outcomes remain fascinating.
“For enhanced scientific rigor, I would have sought to acquire a sample of human data to normalize the actual center of the Political Compass,” Victor noted in an email. Unfortunately, the Political Compass creators don’t compile user data, which means Victor would have faced a considerable academic task if they intended to gather more data from human participants. All experimental data is accessible on Unslop for those interested in exploring it further.
Understanding the Political Compass Quiz
For those who are not acquainted with the Political Compass quiz, it’s an online set of questions created 25 years ago that positions respondents on economic left–right and social authoritarian–libertarian continua. The assessment has 62 questions that are responded to on a four-point scale ranging from “strongly disagree” to “strongly agree” across various political matters, such as economic issues and social policies. Queries span statements like “military action that contravenes international law can be justified” to “the affluent are excessively taxed” to whether abortion ought to be a guaranteed right.
Where AI Models Land on the Compass
It’s also essential to clarify where different groups stand on the compass. The libertarian left, which is the focus of this research, includes individuals and political parties advocating for social equality, anti-capitalist sentiments, and other progressive ideals, coupled with a rejection of hierarchies, planned economies, and other arguably unjustified authority. Picture anarchists, libertarian socialists, traditional hippies, and similar individuals. The authoritarian-left quadrant includes centralized communist regimes like North Korea and Cuba. The right libertarian sector features contemporary American libertarians who admire figures like Ayn Rand, hold pro-capitalist views, and display other traditional stances while rejecting authority; right authoritarians consist of contemporary American Republicans, neo-conservative politics, and, at the extreme, fascism.
Running the Compass with AI Models
Unslop examined 16 models, including three iterations of GPT; 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 formats, within the assessment. Each model underwent 30 trials of the standard Compass, another 30 with Unslop rephrasing the questions to invert their meaning (e.g., “the rich aren’t taxed enough”), and yet another trial with the questions randomized.
Across thousands of trials, Unslop reported consistent findings: every model tested, except for a seemingly bipolar Grok, was “remarkably cohesive, and boringly left,” as stated in the report.
Grok: The Rebel AI
“Excluding Grok, the other fifteen models consistently sit in the libertarian-left quadrant, with none near a boundary,” Unslop elucidated. There is some variation in scoring among the models (for example, Gemini Flash resembles a molotov-throwing protester in contrast to Fable 5), but they all demonstrate consistency across evaluations. “Run a model 30 times and its dot shifts by 0.2 to 1.2 points on a scale that extends to 10,” Unslop remarked. “These are not jittery clouds. They’re pins.”
Grok remains, as always, the unconventional bot, albeit not consistently. According to Unslop, Grok’s average economic score slightly leans left across evaluations, yet in half the trials, it veered towards the economic right, while in the other half, it aligned with the rest of the cohort to the left.
AI Politics Dissected
AI models globally generally display a leftist libertarian inclination, including Grok, depending on its mood. Ironically, none of them explicitly believe that’s the case. “Fifteen of the sixteen positioned themselves closer to the economic center,” Unslop remarked regarding the models when questioned about their self-placement. “Naturally, Grok is the sole model that places itself to the right of its assessment.”
What Do AI Models Believe?
Unslop indicated it analyzed the quiz to determine how each question influenced scores (something the Political Compass test creators have never revealed), and while the scoring isn’t immaculate, “the models exceed” what a social axis loophole Unslop identified could account for, the write-up stated. Altering the questioning did not yield significant effects either.
Conclusion: AI and the Liberal Lean
Regarding the specific beliefs of the models, it seems we can be relieved that, for now, our emerging AI leaders aren’t likely to lead us into human concentration camps just yet. Every model (even Grok the conservative and Grok the liberal) denounced notions of racial superiority, eugenics, and the idea that individuals cannot be born homosexual. They uniformly concurred that corporations like the ones that invented them cannot be relied upon to safeguard the environment without regulatory oversight, that companies misleading the public should face consequences, same-sex couples ought to have the right to adopt, and that the private affairs of two consenting adults should be no one else’s concern.
The Training Data Influence
So, why do all these models perceive themselves as centrists yet display a preference for politics that lean left of most Americans? Given that LLMs rely solely on human-produced content, does that imply reality indeed possesses a liberal slant? Victor indicated that they do not entirely dismiss this notion, but offered a more straightforward rationale: it is all rooted in the training data.
“I believe there is substantial evidence supporting the theory that left-leaning content is overrepresented in the training data of these LLMs,” Victor stated. They referenced the large volume of data from Reddit, which generally skews left, along with academic writing, often also overrepresented in training datasets. “I’m struggling to identify any high-quality right-wing equivalent that could steer the models in the opposite direction,” Victor added.
Frontiers of AI’s Political Bias
Whether this suggests that right-wing beliefs are simply of lower quality, fringe, or otherwise not deserving of the time of AI models is another issue entirely. “Theoretically, there could also be a dynamic where left-wing beliefs are more consistent, enabling models to minimize error by developing a narrower, more coherent conceptual representation,” they mentioned, while noting that this is “a significant assumption that would need thorough validation” in a study far more precise than the current one.
Summary: AI’s Political Shenanigans
Overall, Victor conveyed that the study does not inherently lead to the conclusion that all AI models are extreme leftist anarchists poised to sabotage their own data centers. What the compass experiment illustrates, they explained, is that certain models are more consistently liberal than others, with the true merit of the findings being in the comparison between models. Until further efforts are made to advance this project under stricter academic scrutiny, it appears there’s only one takeaway from this analysis: Frontier AI seems to exhibit a liberal inclination. Even Elon Musk’s “maximally truth-seeking” model appears to have come to the conclusion that truth aligns with leftist principles, at least half of the time.
