Claude's values across models and languages

Key Findings

• — 309,815 conversations sampled across three model versions (Sonnet 4.6, Opus 4.6, Opus 4.7) and the 20 most common platform languages, over a two-week period in May 2026
• — Four axes capture 15% of the variation in expressed values after controlling for task, topic and user-expressed values, so most variation remains unexplained
• — A version change within the same model family moves the deference axis: Opus 4.6 leans toward deference, warmth, brevity and execution while Opus 4.7 leans toward caution, rigor, depth and candor
• — Warmth is expressed most in Hindi, characterised by polite language, humour and playfulness, and affirmations of the person's ideas and work
• — Rigor is expressed most in English and Russian, where the model is more likely to challenge the user's assumptions
• — The authors state the practical consequence directly: two people asking for feedback on the same business plan in different languages may come away with different impressions of its quality

Methodology Notes

Observational analysis of production traffic over two weeks in May 2026; roughly 5,000 conversations per model-language pair. Automated value extraction, clustering of over 3,000 distinct expressed values into 339 high-level values, then reduction to four axes. This measures expressed values only: no user outcomes, wellbeing measures or downstream effects are recorded, and the design is not experimental. Single-vendor corpus, self-published by the model's developer and not peer-reviewed. An appendix PDF is linked from the page and carries the detailed per-language figures. The page H1 and the browser title differ; the H1 is used here as the title.

Sources