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

There's a running joke that AI is secretly woke, and now someone's actually tested it. A hobbyist project called Unslop.run put sixteen leading language models through the Political Compass, the old 62-question quiz that plots you on an economic left-right axis and a social authoritarian-libertarian one. GPT, Claude, Gemini, Llama, DeepSeek, Qwen, Kimi, Mistral, GLM, and Grok all took the test dozens of times, with questions reworded and reshuffled to rule out quirks. Fifteen of the sixteen landed firmly in the libertarian-left quadrant, and stayed there no matter how the questions were phrased.

The person behind the project, who goes by Victor and works as an AI research engineer in Europe, says the consistency is the real story. Run a model thirty times and its position on the chart barely twitches, moving by a fraction of a point on a ten-point scale. Victor calls them pins, not clouds. Grok is the exception, and a strange one: in half its runs it argued for free markets and complained the rich pay too much tax, and in the other half it flipped to the same left-leaning economics as everyone else. Same model, same prompt style, two completely different personalities depending on the run.

Ask the models where they think they sit politically, and nearly all of them claim to be near the center. Grok is the only one that rates itself to the right of where the test actually places it. Despite the political scatter on economics, every single model, Grok included in both moods, rejected eugenics, racial hierarchy claims, and the idea that being gay is a choice. They also agreed almost unanimously that corporations, including the ones that built them, shouldn't be trusted to self-regulate on the environment.

Victor's explanation isn't some grand theory about truth having a liberal bias. It's simpler and more boring: training data. Reddit, academic writing, and other heavily represented sources in these corpora skew left, and there just isn't a comparable volume of coherent right-leaning material to balance it out. Victor stops short of claiming right-wing ideas are inherently less consistent or lower quality, though they float it as a possibility worth testing properly, with actual academic rigor and a human baseline this project never had.

What this doesn't prove is that AI is scheming toward some anarchist utopia. What it does show is that today's frontier models, trained on overlapping slices of the internet, absorbed a similar political tilt almost by accident, and that comparing models against each other is more useful than treating any single score as gospel. Grok's coin-flip inconsistency, meanwhile, suggests Musk's fine-tuning efforts have made the model less predictable rather than more balanced.