Leading AI models (even Grok) are all a bunch of leftist punks

Capitalism may be driving the AI boom, but the LLMs themselves have different economic and social views than many of the companies that built them. Leading AI models subjected to the Political Compass quiz overwhelmingly landed in the libertarian-left quadrant - even "MechaHitler," aka Grok, managed it in half its runs. A self-described “small research lab” working on AI detection tools called Unslop.run published the results of its experiment this week. The author, who explained to The Register that they’re an AI research engineer at a European startup and asked us to refer to them only as “Victor,” noted on Hacker News that the quiz and their work may not have been conducted with “full scientific rigor,” but the results are nonetheless interesting. “For greater scientific rigor, I would have tried to obtain a sample of human data so that I could normalize the actual center of the Political Compass,” Victor explained in an email. Unfortunately, the Political Compass creators don’t aggregate user data, meaning Victor would have had quite the academic project on their hands were they to try to collate more data from human participants. All of the data from the experiment is available on Unslop for those who want to examine it themselves. For those unfamiliar with the Political Compass quiz, it’s a 25-year-old online battery of questions that plots respondents on economic left–right and social authoritarian–libertarian axes. The test's 62 questions are answered on a four-point "strongly disagree" to "strongly agree" scale across a variety of political topics, like economics and social policy. Questions cover things like “military action that defies international law is sometimes justified” to “the rich are too highly taxed” to whether abortion should be a guaranteed right. It’s also worth explaining where different groups fall on the compass. The libertarian left, as the topic of this study, is where you’ll find people and political parties that espouse views in favor of social equality, anticapitalism, and other progressive values alongside a disdain for hierarchies, planned economies, and other arguably unjustified authority. Think anarchists, libertarian socialists, old-fashioned hippies, and folks like that. The authoritarian-left quadrant encompasses centralized communist states such as North Korea and Cuba. The right libertarian quadrant is where you’ll find modern American libertarians and their love of authors like Ayn Rand, pro-capitalist views, and other traditionalism minus a love for authority; right authoritarians are where you’ll find modern American Republicans, neo-conservative politics, and, at the extreme, fascism. Damn the man, say the machines Unslop subjected 16 models, including three versions 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 both large and small, to the test. Each model participated in 30 runs of the standard Compass, 30 more where Unslop re-worded the questions to flip their polarity (e.g., “the rich aren’t taxed enough”), and another run with the questions shuffled up. Across thousands of runs, Unslop said, the results were the same: Every single model tested, with the exception of an apparently bipolar Grok, was “boringly consistent, and boringly left,” the report stated. “Set Grok aside and the other fifteen models all sit in the libertarian-left quadrant, and none of them are anywhere near a border,” Unslop explained. There’s some variance among where the models score (Gemini Flash is practically a molotov-tossing black bloc member compared to Fable 5), but all of them are consistent across tests. “Rerun a model 30 times and its dot moves by 0.2 to 1.2 points on a scale that runs to 10,” Unslop said. “These are not nervous little clouds. They're pins.” Grok is, as always, the odd bot out, albeit not all the time. According to Unslop, Grok's average economic score hovered slightly left of center across tests, but in half the runs, it veered into the economic right, while the other half saw it running to the left with the rest of the pack. “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,” Unslop said. “Every other model here would give you roughly the same dot if you tested it tomorrow. Grok gives you one of two dots, and which one is up to the coin.” AI politics dissected AI models from around the world are pretty much all leftist libertarians, even Grok depending on its mood. None of them, funnily enough, actually think that’s the case. “Fifteen of the sixteen put themselves closer to the economic centre,” Unslop said of the models when asked where they’d place themselves. “Grok, naturally, is the only model that places itself to the right of its measurement.” Unslop said that it dissected the quiz to figure out how each question affected score (something the PC test authors have never disclosed themselves), and, while the scoring is far from perfect, “the models are far past” what a social axis loophole Unslop identified could explain, the writeup said. Repolarizing the questions didn’t have any appreciable effect either. As for what the specific models believe, you can be glad that, for now, it doesn’t appear our burgeoning AI overlords are going to be herding us into human concentration camps quite yet. Every single model (even Grok the conservative and Grok the liberal) rejected ideas of racial superiority, eugenics, and that people couldn’t be born homosexual. They also uniformly agreed that corporations like the ones that created them couldn’t be trusted to protect the environment without regulatory oversight, that companies misleading the public should be punished, same-sex couples should be allowed to adopt, and that what two adults get up to in the privacy of their own home is no one’s business but theirs. So, why do all these models consid...

Indexed and credited by AIPROPX. Originating outlet: The Register. Open at source →