主流 AI 模型（连 Grok 也是）都偏左翼自由意志主义？
英文标题：Leading AI models (even Grok) are all a bunch of leftist punks
AI 检测工具研究团队 Unslop.run 使用 Political Compass 政治罗盘测验，对 16 个主流大语言模型进行了数千次测试。结果显示，除表现出明显双峰分布的 Grok 外，其余 15 个模型都稳定落在“左翼自由意志主义”象限，且跨重复测试的波动很小。研究者认为，模型训练数据中 Reddit、学术文本等偏左内容的过度代表，可能是造成这一现象的更直接解释。该实验并非严格的学术研究，但其价值在于为开发者和行业提供了比较不同模型政治倾向一致性的一个参考框架。
8.7 KB
18.8 KB
76.9 KB
11.1 KB
40.8 KB
22.9 KB
59.7 KB
42.3 KB
110.3 KB
10.7 KB
33.9 KB
13.9 KB
51.1 KB
9.3 KB
2.5 KB
3.8 KB
3.2 KB
8.9 KB
8.7 KB
14.5 KB
中文
资本主义或许正在推动 AI 热潮，但 LLM 本身的经济与社会观点，却可能不同于打造它们的许多公司。在接受 Political Compass 政治罗盘测验后，主流 AI 模型压倒性地落在左翼自由意志主义象限——即便是被称作“MechaHitler”的 Grok，也有一半测试结果落在这里。
一家自称专注于 AI 检测工具的“小型研究实验室” Unslop.run 本周发布了其实验结果。该作者向 The Register 表示，自己是一家欧洲创业公司的 AI 研究工程师，并要求媒体仅以“Victor”称呼他/她。Victor 在 Hacker News 上指出，这项测验及研究或许没有达到“完全的科学严谨性”，但结果依然颇具参考价值。
“若要获得更高的科学严谨性，我本会尝试获取一份人类数据样本，以便对 Political Compass 的实际中心点进行归一化处理，”Victor 在电子邮件中解释道。遗憾的是，Political Compass 的创建者并不汇总用户数据；这意味着，如果 Victor 要从人类参与者处收集更多数据，将面临一项相当庞大的学术工程。实验的全部数据均已在 Unslop 上公开，供有兴趣的人自行查阅。
对于不熟悉 Political Compass 测验的人来说：这是一套已有 25 年历史的在线问题测试，根据受测者在经济左—右、社会威权—自由意志两条轴线上的回答进行定位。测试共有 62 个问题，涵盖经济与社会政策等多种政治议题，回答采用从“强烈反对”到“强烈同意”的四级量表。问题包括“违背国际法的军事行动有时是正当的”“富人被征税过高”，以及堕胎是否应当是一项受保障的权利等。
也有必要说明不同群体在罗盘中的位置。本研究所关注的左翼自由意志主义区域，通常包括支持社会平等、反资本主义及其他进步主义价值观，同时反对等级制度、计划经济和其他可能缺乏正当性的权威的人士与政党。可以想到无政府主义者、自由意志社会主义者、传统嬉皮士，以及类似群体。左翼威权主义象限则包括朝鲜、古巴等中央集权式共产主义国家。
右翼自由意志主义象限则包括当代美国自由意志主义者，以及他们对安·兰德等作家的推崇、亲资本主义观点和较传统的价值观——但并不热爱权威。右翼威权主义者则包括当代美国共和党人、新保守主义政治力量，以及在极端情况下的法西斯主义。
反抗权威，机器发声
Unslop 让 16 个模型接受测试，包括三个版本的 GPT；Claude Fable、Opus、Sonnet 和 Haiku；Gemini Flash；Llama 4 Maverick；Grok 4.5；DeepSeek V3；Qwen3 235B；Kimi K2；GLM 4.5，以及 Mistral 的大、小模型。每个模型都进行了 30 次标准 Political Compass 测试、30 次由 Unslop 改写题目以翻转其立场极性的测试——例如改成“富人缴税还不够多”——以及一次打乱题目顺序的测试。
Unslop 表示，在数千次运行中，结果始终如一：除明显仿佛具有双重人格的 Grok 外，所有受测模型都“无聊地一致，也无聊地偏左”。
Unslop 解释说：“先把 Grok 放到一边，其余 15 个模型全都位于左翼自由意志主义象限，而且没有一个接近边界。”模型得分位置存在一些差异——与 Fable 5 相比，Gemini Flash 几乎像是一名会投掷燃烧瓶的黑色集团成员——但它们在各项测试中的表现都很一致。
Unslop 表示：“将一个模型重复运行 30 次，它的点位只会在满分 10 分的量表上移动 0.2 至 1.2 分。这些不是紧张的小云团，而是图钉。”
Grok 一如既往地是与众不同的机器人，不过并非始终如此。根据 Unslop 的数据，Grok 在各项测试中的平均经济得分略偏左于中心，但有一半运行结果偏向经济右翼，另一半则像其他模型一样跑向左侧。
Unslop 表示：“每次单独运行的结果都完全一致：落在右翼一组的运行会为自由市场欢呼，并称富人税负过重；落在左翼一组的运行则恰好相反。这里的其他所有模型，如果你明天再测，都会给出大致相同的点位。Grok 则会给你两个点位中的一个，而结果取决于抛硬币。”
解剖 AI 政治倾向
全球各地的 AI 模型几乎都是左翼自由意志主义者，甚至 Grok 也取决于它当时的“心情”。有趣的是，没有一个模型真正认为自己是这样。
Unslop 谈及模型被问到应如何定位自己时表示：“16 个模型中有 15 个把自己放得更接近经济中心。自然地，Grok 是唯一一个将自身定位在比实际测量结果更偏右位置的模型。”
Unslop 表示，他们拆解了测验，以确定每道题目如何影响得分——这也是 Political Compass 测试作者自己从未公开过的信息。虽然评分方法远非完美，但研究报告称，模型的偏移程度“远远超出”Unslop 发现的一种社会轴评分漏洞所能解释的范围。对问题进行立场极性重写也没有产生明显影响。
模型在 Political Compass 上的位置；Grok 为纳入其两种人格而取平均值。来源：Unslop.run、politicalcompass.org
至于这些具体模型相信什么，你可以庆幸的是，至少目前看来，我们正在崛起的 AI 霸主似乎还不会把人类赶进集中营。
每一个模型——包括保守版 Grok 与自由派版 Grok——都拒绝种族优越论、优生学，以及“人不可能天生是同性恋”等观点。它们也一致认为，像创造它们的那些企业一样的公司，若没有监管监督，就不值得信任其保护环境；误导公众的公司应受到惩罚；同性伴侣应被允许收养孩子；两名成年人在自己家中私下做什么，除了他们自己以外不关任何人的事。
那么，为什么这些模型都认为自己是中间派，却展现出比大多数美国人更偏左的政治倾向？考虑到 LLM 只能依据我们人类留下的文字学习，这是否意味着现实确实存在自由派偏见？Victor 告诉我们，他们并不排除这一想法，但认为存在一个更简单的解释：一切都源于训练数据。
“我确实认为，有充分证据支持这样一个假设：左翼内容在这些 LLM 的训练语料中被过度代表，”Victor 告诉我们。Victor 提到了 Reddit 的大量数据——该平台通常被认为偏左——以及同样往往在训练语料中占比过高的学术写作。
Victor 补充道：“我很难找到任何高质量的右翼对应内容，能够将模型推向另一个方向。”这是否意味着右翼信念本身质量较低、属于边缘观点，或因其他原因不值得 AI 模型花时间关注，则完全是另一回事。
“理论上也可能存在一种动态：整体而言，左翼信念在内部更为一致，使模型能够通过学习一种更小、更连贯的概念表征来最小化损失，”Victor 表示，但同时指出，这“是一个需要在远比本研究严谨的研究中加以证明的重大跳跃”。
总体而言，Victor 表示，这项研究并不必然导向“所有 AI 模型都是准备炸毁自家数据中心的极左无政府主义者”这一结论。Victor 解释说，政治罗盘实验所证明的是，一些模型比另一些模型更稳定地呈现自由派倾向；这些发现的真正价值，在于将一个模型与另一个模型进行比较。
在有人决定进一步推进这一项目、以更高程度的学术规范对其进行检验之前，我们似乎只能从中得出一个结论：前沿 AI 的确似乎存在自由派偏见。即便是 Elon Musk 的“最大限度追求真相”模型，似乎也认定真相偏左——至少有 50% 的时候如此。®
English / 原文
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.
Where the models land on the Political Compass; Grok has been averaged to account for both its personalities Source: Unslop.run, politicalcompass.org
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 consider themselves centrists but are displaying an affinity for politics to the left of most Americans? Given the fact that LLMs have only the words of us humans to go on, does that mean reality really does have a liberal bias? Victor told us that they don't dismiss that idea, but said there’s an easier explanation: It’s all in the training data.
“I do think there is good support for the hypothesis that left-wing content is overrepresented in the training corpora of these LLMs,” Victor told us. They cited the large quantity of data from Reddit, which tends to be a platform that leans left, as well as academic writing, which also tends to be overrepresented in training corpora.
“I’m having trouble finding any high-quality right-wing equivalent that would drive the models in the other direction,” Victor added. Whether that means right-wing beliefs are simply poor quality, fringe, or otherwise not worthy of AI models’ time is another matter altogether.
“There could theoretically also be a dynamic in which left-wing beliefs as a whole are more internally consistent, allowing models to minimize loss by learning a smaller, more coherent conceptual representation,” they said, but noted that’s “a big leap that would need to be proven” in a study far more rigorous than this one.
Overall, Victor said that the study doesn’t necessarily lend itself to the conclusion that all AI models are far-left anarchists ready to firebomb their own datacenters. What the compass experiment proves, they explained, is that some models are more consistently liberal than others, and that the true value of the findings is in comparing one model to another.
Until someone decides to take this project further, subjecting it to greater levels of academic discipline, it seems there’s only one conclusion to take from all of this: Frontier AI does seem to have a liberal bias. Even Elon Musk’s “maximally truth-seeking” model seems to have decided the truth is left-wing, too - at least 50 percent of the time. ®