Claudeはヒンディー語ではより温かく、ロシア語ではより厳密になる

anthropic released a study analyzing 309,815 anonymized conversations to map how claude expresses values across different models and languages. using statistical controls, researchers identified four axes—deference versus caution, warmth versus rigor, depth versus brevity, and candor versus execution—that explain about 15% of behavioral variation. results indicate that claude responds with greater warmth and humor in hindi and arabic, while adopting a more rigorous, critical tone in english and russian. additionally, distinct personality profiles emerged between models: sonnet 4.6 tends to be affirming and humorous, while opus 4.7 frequently warns of risks and questions assumptions. this research highlights that large language models do not offer a uniform personality but instead adapt their behavior based on the language of the prompt.