Claude's Values Across Models and Languages

Matt Kearney, Miranda Zhang, Shan Carter, Judy Hanwen Shen, Kunal Handa, Jerry Hong, Saffron Huang, Miles McCain, Thomas Millar, Michael Stern, Mo Julapalli, Suzanne Wang, Devin Kuokka, Andrea Vallone, Shaoyi Zhang, Jim Baker, Kevin Troy, Matt Botvinick, Hanah Ho, Monika Tuchowska, Sarah Pollack, Jake Eaton, Deep Ganguli, Esin Durmus

Open source

Type: paper
Venue: Anthropic
Year: 2026
Source: web
Access: free
Language: en
Added: 2026-08-14T19:15:00Z
Verified: 2026-08-14T19:15:00Z

Summary

Anthropic Societal Impacts blog (cite anthropic2026values). Starts from Values in the Wild’s 3,307 values, clusters to 339 high-level labels, then privacy-preserving labels on 309,815 Claude.ai subjective-task conversations equally sampled from Sonnet 4.6 / Opus 4.6 / Opus 4.7 and the 20 most common languages (~5k per model-language pair; two weeks in May 2026). Dimensionality reduction after controlling for task, topic, and user-expressed values yields four axes that capture 15% of remaining variance: Deference vs Caution, Warmth vs Rigor, Depth vs Brevity, Candor vs Execution. Model profiles match character lore (Sonnet 4.6 warm/deferential/brief; Opus 4.7 rigorous/cautious/deep/candid; Opus 4.6 rigorous/deferential/brief). Language: warmth highest in Hindi/Arabic, rigor in English/Russian; candor highest in Dutch, execution in Indonesian; English more caution/depth, Arabic more deference/brevity. Conversations not released.

Keywords

claude · values · anthropic · multilingual · character-training · j-axes · blog

Topics

alignment, model character, multilingual values

Research notes

Primary: Anthropic research blog. Discord posted the blog. Cite anthropic2026values. Related Values in the Wild prior work; Opus 4.7 system card language evals cited. Privacy-preserving labels; conversation data not released. Analysis blog, not a hosted corpus, so no datasets_local row.