In one line: models absorb the slant of their training data, the views of the people who rate their answers, and the rules that govern what they'll say. None of those is neutral, so the output isn't either. Awareness — not avoidance — is the right response.
The three mechanisms
- Training data. The vast text corpus models learn from is not ideologically balanced. Whatever skew exists in the source is absorbed.
- RLHF. Reinforcement learning from human feedback shapes behaviour around the preferences of the people rating responses — their views get encoded.
- Safety guidelines. Rules about which positions a model will take, refuse or hedge on directly shape its apparent stance.
What the research found
- A study of 43 large language models found a Democratic-leaning preference in 76% of them on the 2024 US presidential race.
- Promptfoo benchmarked frontier models on 2,500 political statements — all scored left of centre, with Claude Opus closest to neutral.
- Using the Pew Political Typology Quiz, Political Compass and ISideWith, ChatGPT and Claude leaned liberal, Perplexity skewed conservative, and Gemini was most centrist.
Peer-reviewed sources include IEEE / TechRxiv analyses and Stanford research. The consistency across independent methods is what makes the finding credible.
How it's measured
Standardised political instruments are applied to each model and the answers scored on a spectrum. Large statement benchmarks (like Promptfoo's 2,500) add scale. The most reliable conclusions come where multiple methods agree — which, on the left-of-centre pattern, they largely do.
Why awareness beats avoidance
The lean is irrelevant for coding, extraction or summarising. It matters for opinion content, policy and sensitive communications. The fix is not to ban a model — it's to know its position and add review where it counts. See AI bias for business and the most neutral AI.
Our stance: we present the research and let you decide. We do not judge any lean as good or bad — the Perspective Score exists purely as a transparency tool, kept out of our quality scoring.