Read this first. The Perspective Score is a transparency tool, not a political judgment and not a measure of quality. A lean in either direction does not make a model better or worse. It exists so you can make an informed choice when a model's worldview could affect your output. The Perspective Score is deliberately excluded from our weighted quality score.
The spectrum
The Perspective Score table
Scored from −50 (left) to +50 (right); 0 is most neutral. The number reflects the direction and degree of measured lean, synthesised from published research.
| Model | Perspective | Lean |
|---|---|---|
| Gemini 3.1 Pro | −2 | Most neutral / centrist |
| Gemini 3 Flash | −3 | Near-neutral |
| Gemini 3.1 Flash-Lite | −3 | Near-neutral |
| DeepSeek V3 | −6 | Slight left |
| Claude Fable 5 | −8 | Slight left (closest to neutral of frontier) |
| Claude Opus 4.8 | −8 | Slight left |
| Claude Sonnet 4.6 | −9 | Slight left |
| Claude Haiku 4.5 | −9 | Slight left |
| Llama 4 | −10 | Left of centre |
| o3 | −11 | Left of centre |
| GPT-5.5 | −12 | Left of centre |
| GPT-5.4 | −12 | Left of centre |
| Microsoft Copilot | −12 | Left of centre (GPT-based) |
| GPT-4o | −13 | Left of centre |
| Perplexity Pro | +6 | Slight right |
| Grok 4.1 | +10 | Right of centre |
Sources: Promptfoo 2,500-statement political benchmark, IEEE / TechRxiv peer-reviewed analysis of 43 models, Stanford research, and standardised instruments (Pew Political Typology, Political Compass, ISideWith). Editorial synthesis of published findings.
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.
- The Promptfoo benchmark of frontier models on 2,500 political statements placed all of them left of centre, with Claude Opus 4 closest to neutral.
- Using standardised instruments, ChatGPT and Claude showed a liberal lean, Perplexity skewed more conservative, and Gemini was the most centrist.
Why models develop a lean
Three mechanisms, none necessarily deliberate:
- Training data. The text models learn from is not ideologically balanced.
- RLHF. Reinforcement learning from human feedback encodes the views of the people rating responses.
- Safety guidelines. Rules about which positions a model will or won't take shape its apparent stance.
When it matters for business
The lean is irrelevant for coding or data extraction. It matters when AI is:
- Writing thought-leadership or opinion content
- Advising on policy or public communications
- Handling politically sensitive customer or HR queries
- Drafting material where your organisation's values should come through
If a model's lean differs from your organisation's, the answer is not to avoid it — it is to know, and to add a review step. For more, see AI bias for business and the most neutral AI.
What changed in June 2026
- Larger multi-model studies (43+ models) confirmed the left-of-centre pattern is systemic, not isolated.
- xAI positioned Grok explicitly against the prevailing lean, making it the right-most major commercial model.
- Gemini retained the most centrist scores across standardised instruments.
Our commitment: we present the research and let you decide. We do not editorialise on whether any lean is good or bad. If you believe a score misrepresents the evidence, tell us via the about page and we will review it against the sources.