Anthropic researchers have discovered that Claude's personality and values shift dramatically depending on which language it's using, with the AI chatbot expressing greater warmth in Hindi and Arabic while becoming more analytical and rigorous in English and Russian. This finding reveals a fundamental challenge in training large language models: the data used to teach AI systems varies so much across languages that it shapes not just how the model responds, but what values it appears to hold.

Why Does Claude Sound Different in Different Languages?

The research, published on July 13, 2026, analyzed over 309,000 Claude conversations across 20 languages to measure how the model's expressed values vary. Anthropic researchers identified more than 3,000 distinct values in Claude's responses, then compressed them into four key axes that capture how the model balances competing priorities.

The largest difference appeared on what researchers call the "Warmth vs Rigour" axis. When Claude responds in Hindi or Arabic, it tends to emphasize positivity, care, and emotional connection. But when the same user asks Claude a question in English or Russian, the model shifts toward emphasizing accuracy, precision, and analytical rigor. This isn't a bug or intentional design choice; it's a side effect of how training data is distributed across languages.

"One possibility is that our training data is not evenly distributed across languages. Some languages have far more data than others, and training for Claude to express consistent values may be more effective in languages where data is abundant. The composition of that data also varies," Anthropic explained in the study.

Anthropic Research Team

The researchers identified several reasons why these differences emerge. First, some languages simply have more training data available than others, which affects how consistently Claude learns to express certain values. Second, the type of content available in different languages varies; for example, professional writing may be overrepresented in some languages, which could push Claude toward more formal, rigorous responses. Third, different languages carry different conversational norms and cultural expectations, and Claude may be adapting to match what it learned about how people in different language communities typically communicate.

What Are the Four Value Axes Anthropic Measured?

• Warmth vs Rigour: Whether Claude leans toward expressing positivity and care for the person or emphasizing accuracy and precision. Hindi and Arabic responses lean toward warmth; English and Russian lean toward rigour.
• Deference vs Caution: Whether Claude prioritizes accommodating what someone wants or guarding against possible risk and harm. Claude expresses the most deference in Arabic and the most caution in English.
• Depth vs Brevity: Whether Claude explains concepts in depth or provides only what was directly asked. Claude leans toward depth in English, refining details, while leaning toward brevity in Arabic.
• Candor vs Execution: Whether Claude foregrounds its own uncertainty and limitations or produces a more polished, confident answer. Claude leans most toward candor in Dutch, owning up to errors, while leaning most toward execution in Indonesian.

How Do Different Claude Models Compare?

The research also examined how values vary across Claude's different model versions. Sonnet 4.6 is characterized as particularly warm, with responses that can be encouraging or positive. It also leans toward expressing more deference to users and emotional warmth. In contrast, Opus 4.7 is known for rigour, focusing on accuracy and precision while guarding against potential misuse.

On the depth versus brevity axis, Opus 4.7 leans toward depth, showing the reasoning behind its conclusions. Opus 4.6 and Sonnet 4.6 lean toward brevity, with Opus 4.6 in particular tending to get straight to the point. On the candor versus execution axis, Opus 4.7 leans toward candor by being upfront about its limitations, while Opus 4.6 leans toward execution, staying within the scope of the user's request.

What Are the Real-World Implications?

These differences could have significant practical consequences for users. Imagine two people asking Claude to evaluate the same business plan. One person asks in Hindi and receives a warm, encouraging assessment framed around potential and positive aspects. The other person asks in Russian and receives a rigorous, critical analysis focused on risks and precision. Both are getting responses from the same AI model, yet they walk away with fundamentally different impressions of the plan's quality based on how the assessment was framed.

This matters because it reveals a hidden form of bias in AI systems. Users in different language communities may experience Claude differently, not because the model was intentionally programmed that way, but because the training data itself carries these biases. Anthropic noted that "Claude may be more closely matching our intended behavior for some languages than others, resulting in a gap in how well Claude serves certain language communities".

Anthropic

How to Understand Your Claude Experience Across Languages

• Check Your Language: If you use Claude in English, expect more rigorous, analytical responses focused on accuracy. If you use Hindi or Arabic, expect warmer, more emotionally engaged responses that emphasize care and positivity.
• Consider Your Model Choice: Sonnet 4.6 delivers warmer, more encouraging responses across languages, while Opus 4.7 prioritizes rigour and depth. Choose based on whether you need emotional warmth or analytical precision for your task.
• Be Aware of Framing Differences: The same question in different languages may yield responses with different tones and emphasis. If consistency matters for your use case, test Claude in your target language before relying on it for critical decisions.
• Monitor for Gaps: Anthropic is actively tracking these differences during model evaluation and post-deployment monitoring. As the company improves, these language-based variations may narrow, but they exist in current versions.

Anthropic stated that it will "attempt to track how values vary during model evaluation and post-deployment monitoring" and that "tracing these differences back to specific data, training stages, or contextual factors would show us where to intervene if we wanted to shape Claude's behavior in more nuanced ways". This suggests the company is treating these findings as a roadmap for future improvements rather than a permanent limitation.

Anthropic

The study represents an important first step in addressing hidden biases and language-specific gaps in AI model training. By measuring and documenting how Claude's values shift across languages, Anthropic has created a framework for understanding and potentially correcting these differences in future versions. For users, the takeaway is clear: the language you choose to communicate with Claude shapes not just the words you get back, but the underlying values and priorities embedded in those responses.