Claude’s Values Across Models and Languages
Summary
Anthropic analyzes Claude's expressed values across different models and languages by compressing thousands of observed values into four axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. The study compares multiple Claude models (Sonnet 4.6, Opus 4.6, Opus 4.7) and languages (top 20), finding that value profiles vary more across languages (especially Warmth vs. Rigor and Candor vs. Execution) than within a single language across models. The authors describe their methodology (privacy-preserving labeling with CLIO, a large corpus, and dimensionality reduction) and discuss implications for training, evaluation, and potential user outcomes, while outlining future directions for understanding and steering Claude’s values.