Measuring political bias in Claude

Anthropic open-sources political bias eval methodology for LLMs.
- 1,350 paired prompts across 150 political topics, three dimensions.
“Paired Prompts”: identical question asked from opposing ideological angles.
- Measures even-handedness, counterargument acknowledgment, and refusal rate.
Claude Sonnet 4.5: 94% even-handedness; Opus 4.1: 95%.
- GPT-5: 89%; Llama 4: 66%; Gemini/Grok nominally higher.
Claude refusal rates 3-5% — low; opposing perspectives acknowledged 35-46%.
Training instills character trait: avoid rhetoric that sways or propagandizes.
GitHub release invites industry-wide reproduction and improvement of standard.