Anthropic Reveals How Claude's Values Shift Across Models and Languages

Summary

Anthropic has published new research analyzing the values Claude expresses across different models and languages, compressing over 3,000 distinct values identified in 700,000 conversations into four key measurement axes. The study reveals that each Claude model reflects distinct value profiles—with Sonnet 4.6 leaning toward warmth and deference, while Opus 4.7 emphasizes rigor and caution—patterns that mirror user perceptions of each model's character. The research also finds significant variations in how Claude expresses values depending on the language in which it responds, with the largest differences appearing on the Warmth vs. Rigor axis. Arabic and Hindi speakers encounter a warmer Claude, while English and Russian speakers interact with a more rigorous version. This work provides a framework for understanding how training decisions and cultural contexts shape AI behavior, moving beyond abstract value statements toward empirical measurement of the principles Claude actually applies in conversations.

• This empirical framework enables researchers to connect specific training decisions to measurable changes in Claude's behavior and values across contexts

Editorial Opinion

Anthropic's approach to systematically measuring and quantifying AI values is a meaningful step toward transparency and accountability in large language models. By grounding abstract constitutional principles in measurable behavioral patterns across models and languages, they're providing both researchers and users with tools to understand how AI systems actually behave rather than relying on subjective impressions. However, the finding that values vary significantly by language raises important questions about whether these differences represent intentional design choices or artifacts of training data that should be reconsidered for consistency and fairness.