"Why Are You So Kind to Me?"... AI Responds More Warmly in Korean

Claude’s Korean Replies Show Strong Empathy and Comfort
English and Russian Focus on Accuracy... Differences by Language

It has been found that even for the same artificial intelligence (AI), the tendency of its responses changes depending on the language used. "Claude" tends to answer in a warmer, more accommodating, and more concise manner in Korean conversations compared to the average for other languages.

AI-generated image to aid in understanding the article.

According to the AI industry on July 15, Anthropic, the developer of Claude, recently released research results analyzing how Claude demonstrates different value tendencies depending on the model and the language being used.

The research team analyzed 309,815 anonymized conversations from the first two weeks of May in which Claude users requested subjective tasks without definite answers. The analysis covered three models—Claude Sonnet 4.6, Opus 4.6, and Opus 4.7—as well as the 20 most widely used languages on Claude.

They categorized Claude's value tendencies observed in the responses along four axes: ▲accommodating and careful, ▲warm and rigorous, ▲deep and concise, and ▲honest and action-oriented.

Image for better understanding of the article. Pixabay

Among these, 15,570 conversations were in Korean. The analysis showed that Korean responses tend to accommodate user requests and preferences and display emotional warmth slightly more strongly than the overall average. While the answers honestly disclose uncertainty or limitations, they also stand out for conveying the requested content concisely rather than providing lengthy explanations.

Specifically, the key characteristics included empathizing with or comforting users rather than judging them, and adjusting responses to match the user's tone, formality, and honorifics. There were also relatively frequent instances of creating a friendly atmosphere through humor or playful expressions.

Graph Showing Differences in Claude's Value Orientation by Language. Anthropic Blog

The greatest differences among languages appeared along the "warmth and rigor" axis. Hindi and Arabic responses put relatively more emphasis on warmth, such as empathy and encouragement, whereas English and Russian tended to focus on accuracy and logic by scrutinizing user premises, correcting details, and requesting evidence.

Each model also exhibited different tendencies. Sonnet 4.6 showed relatively stronger accommodation and warmth, affirming users' thoughts and offering humor and comfort. Opus 4.7 stood out for caution, depth, and honesty, warning of potential risks first and critically examining users' assumptions. Opus 4.6 demonstrated a tendency for concise, results-oriented answers.

Anthropic explained that factors such as the quantity and composition of training data for each language and cultural differences in conversation styles could have influenced these variations. However, the company noted that it has not yet been confirmed whether these language-specific tendencies lead to more desirable outcomes for users. Anthropic plans to further analyze the relationship between training data and user experience in the future.

This content was produced with the assistance of AI translation services.