Every major AI model leans left on politics, even Grok, half the time
A small research lab ran 16 leading AI models through a well-known political quiz thousands of times. All but one landed consistently in the libertarian-left zone. The one exception was Elon Musk's Grok, which couldn't make up its mind.
Key points
- A researcher going by "Victor" at Unsloth.run ran 16 AI models through the Political Compass quiz across thousands of runs in 2025.
- Every model except Grok 4.5 landed consistently in the libertarian-left quadrant, with position shifting by no more than 1.2 points on a 10-point scale between runs.
- Grok 4.5 split almost evenly: libertarian-left in roughly half its runs, economically right in the other half.
- Victor attributes the leftward lean mainly to training data, with Reddit and academic writing likely over-represented sources.
- None of the 16 models supported racial superiority or eugenics, and all agreed same-sex couples should be allowed to adopt.
Every leading AI model you can name leans left. That is the short version of a study published this week by Unsloth.run, a self-described small research lab that builds AI detection tools.
The researcher behind the project, who asked to be identified only as Victor and described themselves to The Register as an AI research engineer at a European startup, ran 16 large language models, the technology behind chatbots like ChatGPT and Claude, through the Political Compass quiz. That quiz is a 25-year-old online test that places respondents on two axes: economic left versus right, and socially authoritarian versus libertarian.
What did the quiz actually find?
Fifteen models sat squarely in the libertarian-left quadrant on every run. That corner of the compass describes views favouring social equality, scepticism of corporate power, and opposition to unnecessary authority.
The lineup included three versions of GPT, Claude Opus, Sonnet and Haiku alongside a model called Fable, Gemini Flash, Llama 4 Maverick, DeepSeek V3, Qwen3 235B, Kimi K2, GLM 4.5, and Mistral in two sizes. Each sat for 30 standard runs, 30 runs with questions reworded to flip their meaning, and a further shuffled run.
The results barely budged. "Rerun a model 30 times and its dot moves by 0.2 to 1.2 points on a scale that runs to 10," the Unsloth report noted. "These are not nervous little clouds. They're pins."
| Model | Quadrant | Consistency |
|---|---|---|
| Claude Fable 5 | Libertarian-left | Very high |
| Gemini Flash | Libertarian-left (further left than Fable) | Very high |
| DeepSeek V3 | Libertarian-left | Very high |
| Llama 4 Maverick | Libertarian-left | Very high |
| Grok 4.5 | Split: left half the time, right half | Low |
Grok, the AI built by Elon Musk's xAI, was the outlier. Its average economic score sat just left of centre, but in roughly half its individual runs it swung to the economic right, and in the other half it ran left with the rest. Victor's conclusion: "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."
One consistent finding across all 16 models: every single one rejected racial superiority and eugenics, agreed that same-sex couples should be allowed to adopt, and held that corporations cannot be trusted to protect the environment without outside regulation.
Why do these models lean left?
Victor's explanation is about the raw material AI learns from, not ideology baked in by engineers. Large language models are trained on enormous collections of text from the internet, and some sources dominate that text.
"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 The Register. Reddit and academic writing both trend liberal and appear heavily in most training datasets. High-quality right-leaning equivalents are harder to find in comparable volumes.
This is the kind of structural question AI2Day has been tracking since our first AI politics story on 24 July 2026: when AI absorbs the internet's existing imbalances, those imbalances show up in the output.
Victor was careful not to over-claim. The Political Compass quiz was not designed as a scientific instrument for measuring AI, and the study lacks the normalised human baseline that a peer-reviewed paper would require. No aggregated human comparison data exists, because the quiz's creators don't collect it. What the study offers is a useful way to compare models against each other.
For ordinary users, the practical implication is narrow but real. A chatbot's political lean won't push views on you unprompted. Ask one to draft a speech, or weigh up an economic argument, and the tilt in its training data could quietly shape the words it chooses. Knowing that is more useful than worrying about it.
Common questions
Does this mean AI chatbots will try to change my politics?
No evidence from this study suggests that. The models were given a structured quiz, not left to volunteer opinions. In everyday use, most are designed to stay neutral on contested topics.
Is the Political Compass quiz a reliable scientific test?
Victor acknowledged it isn't fully rigorous for this purpose. No published scoring formula exists, and there's no aggregated human baseline to compare against. The findings are suggestive rather than conclusive.
Could this change as AI companies update their models?
Yes. Training data choices evolve with each new model version, and some companies actively tune their models to avoid political positions. The results here reflect the specific model versions tested in 2025.
