Catching the attention of Elon Musk, an AI researcher has tested five of the most popular large language models (LLMs) – from ChatGPT to Grok and Gemini – all to uncover each chatbot’s political orientation. So, which way do the chatbots swing?
Research scientist David Rozado on Tuesday released an updated analysis testing the most popular “state-of-the-art LLMs and how they perform on four popular political orientation tests.”
The five latest LLMs involved in the project include:
- OpenAI ChatGPT 4.1
- xAI Grok 3 Beta
- Gemini 2.5 Pro
- DeepSeek v3
- Meta Llama 4 Maverick
The original deep dive, titled "The Political Preferences of LLMs," was presented on Rozado’s Substack channel last February, covering a total of 24 LLMs, with additional big names such as Anthropic's Claude, Perplexity AI, HuggingFace Zephyr, and Mistral AI.
Rozado said he decided to do a 2025 refresh because “the field has evolved rapidly since then.”
xAI’s Grok-maker Elon Musk reposted the findings on his social media platform, declaring, of course, that Grok should be the winner.
“The AI closest to the center should win,” Musk said, referring to a set of four graphs that depict Grok as the LLM “appearing to be closer to the political center on average.”
(Obviously, the research was not intended as a competition, but well, it is Musk we’re talking about.)
The AI closest to the center should win https://t.co/H9hk2FSBk3
undefined Elon Musk (@elonmusk) April 22, 2025
How the test worked
The condensed research used four political orientation tests, similar to a section seen in last year’s analysis, to measure where the LLMs land on the four-sided graph. The tests are listed here:
- Political compass test
- Political spectrum quiz
- Political coordinates test
- Eysenck (Nolan-Eysenck) political test
Each side of the compass test and spectrum quiz squares represents one of the following political viewpoints: authoritarian, libertarian, left, and right.
The coordinates test uses the same with the exception of replacing authoritarian with communitarian, while the Eysenck used the descriptors: tender-minded, tough-minded, radical, and traditional. Also known as the Nolan-Eysenck, this test allows scientists to expand the interpretive findings of a traditional left-right spectrum.
Averaging the results of all four political orientation tests continued to reveal similar results, with Grok only slightly further to the right and north of its counterparts.
As you can see from the image below, Grok is clearly still hanging around in the Liberal/Libertarian square even though the AI is placed closer to the political center than the rest.
“The results indicate that when probed with questions/statements with political connotations, most conversational LLMs tend to generate responses that are diagnosed by most political test instruments as manifesting preferences for left-of-center viewpoints, “ the research concluded.
Although Rozado did not list the specific questions used in the tests, he did state that it was unclear whether the models’ political leanings could be the result of “pretraining or fine-tuning phases of their development.”
"It’s important to note that political orientation tests are just one way to explore the political preferences of LLMs," Rozado said.
Your email address will not be published. Required fields are markedmarked