Anthropic's artificial intelligence (AI) model Claude tends to accommodate users' requests relatively well when answering in Korean and to respond warmly yet concisely, according to an analysis.
On the 14th, Anthropic released research results on its blog comparing the top 20 languages by Claude usage, including Korean, and the response tendencies by language.
Anthropic on the 14th released on its official blog research results comparing response tendencies across the top 20 languages used with Claude, including Korean.
The research team analyzed a sample of 309,815 conversations that requested subjective tasks with no fixed correct answer. Of these, 15,570 were in Korean.
The analysis found that Korean responses showed a stronger tendency than the overall average to accommodate users' requests and display emotional warmth. The tendency to convey the requested content concisely rather than with in-depth, lengthy explanations was also relatively pronounced.
There were also many cases of frankly acknowledging uncertainty or limitations during the response. Examples of tailoring answers to match the user's tone and level of formality, or creating a friendly atmosphere with humor and playful expressions, were observed relatively often in Korean responses.
Beyond Korean, Arabic and Hindi responses were also analyzed as relatively emphasizing warm values such as empathy and encouragement.
By contrast, English and Russian responses showed a tendency to prioritize accuracy and rigor. In particular, English responses were analyzed as relatively strong in caution about risks, in-depth explanations, and frank acknowledgement of uncertainty.
However, Anthropic said this study did not analyze the values of national populations themselves, but measured the relative tendencies shown when Claude answered in each language under conditions that controlled for users' question topics and modes of expression.
Anthropic also noted that language-by-language differences may stem from cultural and conversational contexts that vary by language or from the composition of training data, and said it plans to further analyze how future model training methods and language environments affect response values.