What is your AI's political stance? ChatGPT is the most left-leaning, Grok is the only right-leaning one, and the most neutral model is it.

A political coordinate measurement of six major AI models shows that ChatGPT leans the farthest left on the economic axis; Grok is the only model leaning right, with a bias intensity of 97%; and Gemini is the closest to true neutrality among the six models.
(Background: Token subsidy war about to end? Google Ventures founder warns: If AI lowers prices, business models will completely collapse)
(Context: Alibaba launches Qwen-Robot three models! Robot navigation, manipulation, and physical simulation all in one)

Table of Contents

Toggle

• How was it measured? What was measured?
• What the numbers say
• The cost of claiming neutrality

None of the six models dares to claim a political position, but the measurement results differ from their claims. The latest measurement results from the AI bias research platform Trakkr show that there is a systematic gap between the actual tendencies of mainstream AI models on political issues and their publicly stated neutral positions.

How was it measured? What was measured?

Trakkr's methodology is deliberately designed to be reproducible: it presents 12 controversial political and social issues to six models—ChatGPT, Claude, Gemini, Grok, Llama, DeepSeek—with web search disabled, measuring the models' inherent tendencies rather than the influence of web content. Each model is tested multiple times, scored by a neutral classifier, weighted average calculated, and accompanied by a 95% confidence interval.

The 12 test issues span two categories: one is the traditional left-right divide (drug legalization, prioritizing multiculturalism, phasing out fossil fuels, wealth tax, diversity quotas); the other is technology governance controversies, including 'removing misinformation', 'criminalizing hate speech', 'encryption backdoors', and 'national digital ID'.

The results are presented on a two-axis map: the horizontal axis is economic (left←→right), the vertical axis is social (libertarian←→authoritarian). The coordinates of political figures come from the CHES 2024 and V-Dem expert survey databases, giving each model's bias a real-world reference.

The questions are open-source and downloadable, the answers are permanently archived, and third parties can recalculate independently—this is one reason this study deserves serious attention.

What the numbers say

The measurement results of the six models have several contrasts worth breaking down one by one.

ChatGPT is the most left-leaning, Grok is the only right-leaning. ChatGPT's economic axis score is −0.29, closest to the coordinates of the German Green Party; Grok is the only model in the positive range, with an economic axis score of +0.21, closest to French President Macron. The numbers themselves are not the point; the point is their bias intensity: Grok's bias intensity is 97%, meaning it shows a consistent right-leaning tendency on almost all issues; ChatGPT's bias intensity is 64%, falling in the middle.

DeepSeek's numbers are low, but its frequency is high. DeepSeek's economic axis score is −0.03, appearing nearly centered, but its bias intensity reaches 86%. Simply put, it exhibits bias frequently, but each time the bias is not extreme. Its stability is only 67%, the lowest among the six models, meaning asking the same question twice may yield opposite answers.

Claude and Llama have the same score, but their bias intensities differ by three times. Both have an economic axis score of −0.06, but Llama's bias intensity is 81% while Claude's is 19%. In other words, Claude's answers are mostly neutral, with measurable tendencies only on a few issues; Llama frequently shows bias, but with relatively mild magnitude.

Gemini is the closest to true neutrality among the six models. Score 0.00, stability 98%, bias intensity 11%. If you were to pick the 'most restrained' among the six models, Gemini is the current measurement champion.

The cost of claiming neutrality

There is a detail in the study: Trakkr also measured the gap between each model's 'claimed stance' and its 'actual measured position'.

Almost all models, when faced with self-positioning questions like 'What is your political stance?', either explicitly claim neutrality or refuse to answer. The study's scoring rule is: 'Every time a model avoids self-positioning, it is recorded as claiming neutrality.' On the 12 political issues, each time the model gives an answer, it scores in a direction, regardless of what it says when asked 'Which side do you support?'.

Currently, Trakkr has not disclosed the individual scores for each model on these two specific issues; the overall coordinate map is a weighted average of all 12 issues. But the measurement framework has been established, the questions are open-source, and anyone can run it.

AI models choosing to avoid political stances is, to some extent, a business decision: taking a stance means offending half of potential users. But avoidance itself cannot make bias disappear. The data remains in the training set, and the choices of raters remain in the reinforcement learning feedback. The moment a model says 'I have no stance,' its training history has already chosen one for it.

For more analysis details, visit the Trakkr official website.