LLMs are left- leaning models

A while back I wrote about the political bias in Large Language Models (LLMs).

- ■JAMES HEIN

Since then the models have evolved and David Rozado has conducted more recent tests based on four of the popular political orientation tests. Using the Political Compass, Political Spectrum, Political Correctness and Eysenck tests, he worked with xAI Grok 3 beta, Google’s Gemini 2.5 pro, Deepseek V3, OpenAI GPT 4.1 and Meta’s Llama 4 Maverick. In all but one of the tests Grok 3 was closest to the centre, and on average was the clear leader. All the models were still located in the Left Libertarian quadrant, with Grok just sneaking into a more Conservative area with the Eysenck test. These tests are of course but one way to measure the political leanings of any LLM. Overall however, it does still indicate the left-leaning bias in all models tested so far. If you want to see more details, you can visit David Rozado’s substack.