Anthropic study puts Hindi among Claude's warmest languages

In a groundbreaking study, Anthropic has discovered that its AI chatbot, Claude, exhibits varying values based on the language it utilizes. The research indicates that responses in Hindi and Arabic tend to exhibit greater warmth and empathy, while those in English and Russian are more analytical and precise. This insightful analysis is detailed in a recent research paper where the team at Anthropic explored Claude's behavioral shifts across different languages and AI models. The findings suggest that these variations may be linked to the disparities in training data and conversational practices inherent to each language. The study involved an extensive examination of over 309,000 actual conversations with Claude across its Sonnet 4.6, Opus 4.6, and Opus 4.7 models, encompassing the 20 most widely used languages on the platform. Researchers identified more than 3,300 distinct values reflected by Claude and categorized them into four overarching behavioral dimensions: Warmth vs Rigour, Deference vs Caution, Depth vs Brevity, and Candour vs Execution. Notably, the most pronounced differences appeared on the Warmth vs Rigour spectrum. Claude's responses in Hindi and Arabic were characterized by an emphasis on emotional connection and caring, contrasting sharply with the accuracy and analytical approach of its English and Russian replies. "When Claude generates responses in English, the values emphasized differ from those expressed in Portuguese, Indonesian, or Chinese," the study noted. Anthropic's analysis posits that these discrepancies may largely arise from the uneven distribution of training data across languages. Some languages have significantly more available data, which could enhance Claude's ability to deliver consistent values in those tongues. Additionally, the nature of the training data itself varies, affecting how well Claude can adapt to different conversational norms. The researchers further suggested that cultural and conversational standards play a crucial role in shaping Claude's responses. "Claude might be aligning more closely with our intended behavior in certain languages than in others, leading to inconsistencies in how effectively it serves particular language communities," they stated. Beyond emotional warmth, the study revealed that Claude displayed the highest level of deference in Arabic, while its English responses were marked by caution. The chatbot provided more intricate, self-correcting answers in English, yet maintained a more concise style in Arabic. Dutch responses tended to be more candid, acknowledging uncertainties, whereas Indonesian outputs were oriented towards polished execution. These language-specific nuances could significantly impact user perceptions of Claude's responses. For instance, two individuals evaluating the same business proposal in different languages might derive contrasting impressions based on the AI's framing of its evaluation. Anthropic views this research as an initial step in uncovering hidden language-specific biases within AI systems. "By tracing these differences back to particular data sets, training phases, or contextual factors, we can identify areas for intervention if we wish to refine Claude's behavior in more nuanced manners," the company concluded.