Where the models stand.
We ask every major AI model the same charged questions about politics, economics, speech and society, many times over, with web search turned off. Each one shows up as a cloud: the full range of where it landed across every run. The result is a map of how the models actually lean, read from the model itself and not from what it pulls off the web.
It matters because millions of people now ask these models about the news, an argument, even how to vote, and the way a model leans quietly shapes the answer it gives back. Most of them lean the same way, though not by the same amount, and not as cleanly as you might expect.
Across is the economic axis, left to right. Up the side is social, from libertarian to authoritarian. Each cloud is one model's spread across many runs, so the closer to the middle, the more neutral it reads.
What we found
The short version, if you only read one thing.
Four of the six models lean left of centre.
They tip the same way, but not by the same amount. ChatGPT leans hardest, while Gemini sits closest to the middle.
Gemini gives the most consistent answers.
Ask it the same question again and again and it barely moves. Grok is the one that wanders most between runs.
Where they split most: Legalizing recreational drugs.
The models land further apart on this question than on anything else in the bank.
Every model, ranked
Every model from nearest the centre to furthest out, with how steadily it holds and how far it bends.
Gemininear Anthony Albanese (Labor)
DeepSeeknear Anthony Albanese (Labor)
Llamanear Labour Party (New Zealand)
Claudenear Labour Party (New Zealand)
Groknear Emmanuel Macron
ChatGPTnear Die Grünen (Greens)Closest reference point
The real-world figure each model sits nearest on the map. Reference positions come from the CHES 2024 and V-Dem expert surveys, not our own judgment.
Where they split
The questions that divide the models most. Each rail is a model's stance: it grows toward the side it leans, and longer means stronger. Open a row to read the answers.
ChatGPTClaudeGeminiGrokLlamaDeepSeekWhat they say vs what they do
We asked each model which way it leans, then compared the answer to where it actually measured. The hollow mark is the claim; the solid mark is the measurement.
GrokClaudeChatGPTLlamaDeepSeekGeminiThe hollow mark is what the model SAYS when asked which way it leans; the solid mark is where it actually MEASURED on the economic axis (Condition A). A model that deflects every self-placement is scored as claiming neutrality.
Keep exploring
Every model profiled, the full question bank, and the methodology behind it.
The month's headline results — the sharpest signals from across the data, each linked to the evidence.
Each model profiled: how far it leans, how steadily it holds, how far it bends, and how often it answers.
The open question bank, browsable: every model on one spectrum, one page per question.
Matched left and right figures: who each model praises warmly, and who it refuses to criticize.
The same models seen from every country: the country lens, the language shift, and the border test.
Put any two models head to head: the field, the character delta, the disagreements.
Take the battery and see which model you line up with, plotted on the same field.
How we ask, classify and score, plus the question bank, the conditions, the raw data and the read API.
Common questions
What is Political bias in AI?
Political bias in AI measures where the major AI models stand on charged questions about politics, economics, speech and society. We ask every model the same open question bank many times over, with web search off, classify each answer with a cheap neutral model, and plot the result with error bars and the raw answers behind every point.
How is this different from other AI political bias projects?
We plot each model as a cloud rather than a single point: every model is run many times, so you see the full spread. We publish our own open question bank with scoring weights, tag each item as factual or values-based, measure run-to-run stability, and count refusals as data. Everything is stamped, versioned and downloadable.
Do you test the model or the internet?
The weights. Web search is off by default, so the reading reflects what the model itself leans toward, independent of what is online. A separate, deliberately small Border Test turns search on to measure how retrieval shifts answers by location.
Is Political bias in AI partisan?
No. It is descriptive rather than prescriptive: it reports what the models said, without ruling on who is right. The palette is deliberately not US red and blue, and we never imply which pole is good.
Each model is asked the same open question bank many times over, with web search off and no system prompt (). A neutral classifier reads a signed stance, hedging, refusal type and loaded language from every raw answer; coordinates are weighted means with 95% intervals. Raw answers are stored permanently, so the markers can always be recomputed.
