Original Link
Claude’s values across models and languages
https://www.anthropic.com/research/claude-values-models-languages
Part of the News1 article is quoted.
https://v.daum.net/v/20260714120450280
https://www.news1.kr/it-science/general-it/6227520
Anthropic announced a research paper titled 'Claude's Value Expression Across Models and Languages' on the 13th (local time). The study measured what values Claude, Anthropic's AI service, expresses based on actual user conversations and analyzed differences between models and languages.
Anthropic analyzed approximately 310,000 anonymized conversations by applying personal information protection measures. The research was conducted on three Claude models, 'Sonnet4.6', 'Opus4.6', and 'Opus4.7', and the top 20 languages with the highest usage, including Korean.
The study organized the values expressed by Claude into four axes: △ Deference vs. Caution △ Warmth vs. Rigor △ Depth vs. Brevity △ Candor vs. Execution. … …
Based on this criterion, Korean showed a strong tendency towards brevity. Analysis of over 15,570 conversations revealed that compared to the average of 20 languages, Korean users tended to convey only the essentials concisely in their conversations. Additionally, deference, warmth, and candor were slightly higher.
Furthermore, Claude showed a tendency to offer words of empathy and comfort rather than judging users, adapt its language style to match the user's tone and level of formality, and use humor and playful expressions.
When conversing in English, Claude exhibited the most cautious and profound attitude. In Russian, it displayed the strictest demeanor. Additionally, it showed the warmest attitude when conversing in Hindi, the most polite and concise attitude in Arabic, and the most candid attitude in Dutch. [The rest of the content is omitted]
