A new study by AI company Anthropic has found that its chatbot Claude expresses different communication styles depending on the language it is using. According to the research, Claude’s responses in Hindi and Arabic tend to be warmer and more empathetic, while its replies in English and Russian are generally more analytical, precise, and fact-focused.
The findings are part of a research paper published by Anthropic, which explored how Claude’s behaviour varies across different languages and AI models. The company believes these differences are influenced by variations in training data as well as the conversational and cultural norms associated with each language.
For the study, Anthropic analysed more than 309,000 real-world conversations involving Claude across its Sonnet 4.6, Opus 4.6, and Opus 4.7 models. The research covered the 20 most frequently used languages on the platform.
Researchers identified over 3,300 distinct values expressed by Claude and grouped them into four key behavioural dimensions: Warmth vs. Rigour, Deference vs. Caution, Depth vs. Brevity, and Candour vs. Execution.
The most noticeable variation appeared in the Warmth vs. Rigour category. Claude’s Hindi and Arabic responses were more emotionally supportive and compassionate, whereas English and Russian responses focused more on accuracy, logical reasoning, and detailed analysis.
Anthropic said these differences are likely linked to the uneven distribution of multilingual training data. Some languages have significantly larger datasets than others, making it easier for the AI to learn consistent behaviours in those languages. The nature and quality of the available data also differ across languages, which may further influence the chatbot’s responses.
The company also pointed out that cultural communication styles play an important role. Different languages naturally encourage different conversational norms, and Claude appears to adapt its responses accordingly. This may result in the chatbot expressing different values depending on the language in which a user interacts with it.
The research uncovered several other interesting patterns. Claude was found to be most deferential in Arabic and most cautious in English. English responses were generally longer, more detailed, and more likely to include self-corrections, while Arabic replies tended to be shorter and more concise. Dutch responses were the most transparent about uncertainty, whereas Indonesian responses focused more on delivering polished, action-oriented answers.
Anthropic noted that these language-specific differences could influence how users interpret AI-generated responses. For example, two people reviewing the same proposal in different languages may reach different conclusions simply because the AI presents its analysis differently.
The company described the study as an important first step in identifying hidden language-specific biases in AI systems. Anthropic said understanding how training data, language, and cultural context shape AI behaviour could help developers build models that provide more consistent and equitable experiences for users across different languages.
Recent Random Post:
