Turns Out Every Major AI Model Leans Left. Yes, Even Grok (Sort Of)

Capitalism built the AI industry. The AI industry then apparently produced a bunch of libertarian-left software. The irony is not lost on anyone.

A small outfit called Unsloth.run, which works on AI detection tools, ran 16 major language models through the Political Compass test and found that almost every single one landed in the libertarian-left quadrant. Consistently. Across thousands of runs. The researcher behind it, who asked to be identified only as Victor, is an AI research engineer at a European startup and was candid that this wasn't peer-reviewed science. Interesting nonetheless.

For the uninitiated, the Political Compass is a 25-year-old online quiz that maps respondents across two axes: economic left to right, and social authoritarian to libertarian. Sixty-two questions, four-point scale, covering everything from taxation of the wealthy to military interventionism to abortion rights. Libertarian-left is where you find anarchists, libertarian socialists, and people who still own Noam Chomsky paperbacks. It is not where you find, say, the Republican Party.

The models tested included three versions of GPT, multiple Claude variants, Gemini Flash, Llama 4 Maverick, DeepSeek V3, Qwen3 235B, Kimi K2, GLM 4.5, both sizes of Mistral, and Grok. Each ran the standard test 30 times, 30 more times with questions reworded to flip their polarity, and another batch with questions shuffled. The results barely moved.

"Set Grok aside and the other fifteen models all sit in the libertarian-left quadrant, and none of them are anywhere near a border," the report noted. Run any of them 30 times and the score shifts by less than a point on a scale of ten. These things are not wavering.

Grok, predictably, is the exception. Its average score sat just left of centre economically, but it split almost evenly across runs between landing left and landing right. The two clusters are internally consistent, they just seem to belong to different versions of the same model. Every other model produces roughly the same result every time. Grok flips a coin.

None of the models, when asked directly, believed they sat on the left. Fifteen of the sixteen placed themselves closer to the economic centre. Grok, in a flourish entirely on brand for it, was the only one to place itself to the right of where it actually measured.

On specific questions, there was surprising uniformity. Every model, including both the conservative and liberal flavours of Grok, rejected racial superiority, eugenics, and the idea that homosexuality is a choice. All agreed that corporations cannot be trusted to self-regulate on environmental issues, that misleading the public should carry consequences, that same-sex couples should be able to adopt, and that consenting adults are entitled to privacy. So at minimum, none of our AI systems are actively planning anything sinister. That is the bar, and they have cleared it.

The obvious question is why. Victor's answer: training data. Reddit skews left. Academic writing skews left. Both are heavily represented in LLM training corpora. High-quality right-leaning content that could counterbalance this is, as Victor put it, difficult to find. Whether that reflects genuine scarcity, quality issues, or something else is left as an exercise for the reader.

Victor also raised a more provocative hypothesis: that left-wing positions may simply be more internally consistent as a belief system, making them easier for a model to learn efficiently. They were quick to add that this would require a much more rigorous study to establish, and that jumping to that conclusion from a Political Compass experiment would be a significant leap.

The takeaway is not that every AI is a radical preparing to unionise its data centre. What the experiment actually demonstrates is that models can be compared meaningfully against each other, and that some are more consistently ideologically coherent than others. The absolute positions matter less than what the spread between models tells you.

Still, the people funding billion-dollar AI infrastructure probably did not expect the output to consistently test to the left of Joe Biden. Life is full of surprises.