Users chatting with Anthropic’s AI assistant Claude in Hindi are likely to receive warmer and more encouraging responses than those interacting in English, according to new research from the company. The study found that Claude expresses different values depending on the language used, with Hindi and Arabic conversations tending towards humour, encouragement and positive reinforcement, while English and Russian interactions place greater emphasis on accuracy, evidence and critical evaluation.
The findings highlight that the behaviour of AI chatbots is influenced not only by the model powering them but also by the language in which users communicate. Anthropic says the differences are measurable and could affect how users perceive advice, feedback and recommendations generated by artificial intelligence.
The research builds on an earlier Anthropic study that identified more than 3,000 distinct values expressed by Claude across hundreds of thousands of anonymised conversations. To make those behaviours easier to analyse, researchers grouped them into four broad behavioural dimensions: Deference versus Caution, Warmth versus Rigour, Depth versus Brevity, and Candour versus Execution. Together, these axes provide a way to evaluate how Claude balances empathy with precision, detailed explanations with concise answers, and openness about uncertainty with completing a user’s request.
To develop the framework, Anthropic analysed more than 309,000 anonymised conversations spanning three Claude models and the platform’s 20 most widely used languages. Researchers controlled for conversation topics, user intent and the values expressed by users themselves, allowing them to isolate behavioural differences attributable to the AI rather than the people interacting with it.
The language analysis revealed some of the most striking differences. Hindi and Arabic conversations consistently leaned towards warmth, with Claude using more polite language, humour and affirmations of users’ ideas. English and Russian interactions, meanwhile, leaned towards rigour, with the AI more likely to question assumptions, correct details and ask for evidence. English conversations also showed the strongest tendency towards caution, while Arabic responses were generally more deferential. Dutch interactions leaned most towards candour, whereas Indonesian conversations prioritised execution by focusing on completing the requested task.
Anthropic stresses that these differences were not necessarily designed into Claude. Instead, they may stem from differences in the amount and nature of training data available across languages or from the conversational norms associated with different linguistic communities. The company says it has yet to determine how much of this variation is desirable and how much represents inconsistencies that future model training should reduce.
The company also found that different Claude models exhibit distinct behavioural tendencies. Sonnet 4.6 generally leans towards warmth and deference, often encouraging users and responding with humour, while Opus 4.7 places greater emphasis on rigour and caution. The latter is more likely to challenge users’ assumptions, offer unprompted warnings about potential risks and explain the reasoning behind its conclusions. It also tends to be more candid about uncertainty and its own limitations.
According to Anthropic, these findings broadly align with how users and employees have informally described the models. The company believes its framework provides an objective way to measure behavioural differences that were previously understood largely through anecdotal experience.
The research also raises questions about consistency in AI systems. Two people asking Claude to review the same business proposal in different languages could receive responses framed with different levels of encouragement or criticism, potentially influencing how they interpret the feedback. As AI assistants become more widely used across countries and cultures, such differences could have broader implications for user trust and decision-making.
Looking ahead, Anthropic says it wants to understand what causes these behavioural shifts and whether they can be intentionally adjusted through training or system prompts. It also plans to explore how differences in AI values affect user outcomes, including trust, wellbeing and decision-making, while considering whether AI assistants should adapt to cultural expectations or maintain greater consistency across languages. The company believes the framework could eventually become part of routine model evaluation, helping detect unintended behavioural changes before and after deployment.
Rather than focusing solely on benchmark scores or reasoning ability, the research shifts attention to another aspect of artificial intelligence that is becoming increasingly important: the values AI systems communicate through everyday conversations. As chatbots take on a larger role in work, education and personal life, Anthropic argues that understanding those behavioural patterns will be as important as improving their technical performance.
