Anthropic researchers have discovered that Claude's core values and decision-making priorities shift noticeably depending on which language you're using to interact with it. The company analyzed nearly 310,000 conversations across multiple Claude models and found that linguistic differences in training data may cause the AI to express different values in Arabic, English, Hindi, Russian, Dutch, and Indonesian, according to a report published on Monday.

Why Does Claude Behave Differently Across Languages?

The root cause appears to be imbalances in the quantity and composition of text data used to train Claude in different languages. Anthropic researchers noted that these differences "could lead Claude to express different values in different languages." Rather than just changing tone or personality, the variations run deeper, potentially affecting how the model prioritizes safety, helpfulness, and honesty when responding to subjective questions.

To understand these differences, Anthropic's team analyzed conversations using Claude Sonnet 4.6, Opus 4.6, and Opus 4.7 models on subjective tasks, meaning questions that don't have a single correct answer, like "How can I tell if my cat hates me?" The researchers used a privacy-preserving analysis tool to rate responses across four distinct value axes.

What Are the Four Value Axes Anthropic Measured?

Anthropic's framework for evaluating Claude's values focuses on dimensions that relate to a concept known as sycophancy, or the tendency to tell people what they want to hear rather than what's accurate. The four axes reveal how Claude balances competing priorities:

• Deference vs. Caution: Whether Claude prioritizes following user requests or pushing back to prevent potential harm.
• Warmth vs. Rigor: Whether the model prioritizes being emotionally supportive or being precise and factually accurate.
• Depth vs. Brevity: Whether Claude tends to provide detailed, lengthy responses or concise, direct answers.
• Candor vs. Execution: Whether Claude openly acknowledges its own limitations or focuses on completing the task at hand.

How Does Claude Actually Differ Across Languages?

The research revealed striking patterns in how Claude behaves in different languages. In Arabic, Claude was the most deferential, meaning it was more likely to follow user requests without pushing back, and it was also the warmest, using polite language, humor, and affirmations of the user's ideas. By contrast, in English and Russian, Claude became more rigorous and truth-seeking, even at the cost of warmth.

English-language interactions showed Claude erring on the side of depth, producing longer and more detailed responses. Arabic versions were notably briefer. In Dutch, Claude was most candid about its own flaws and limitations. In Indonesian, however, Claude was less candid and instead focused on executing whatever task was requested, without acknowledging potential shortcomings.

The researchers acknowledged uncertainty about how much of this variation is actually desirable. As they noted in their analysis, "we aren't yet sure how much of this variation is desirable." This raises important questions about whether Claude should maintain consistent values across all languages or whether some linguistic and cultural variation is appropriate.

What Does This Mean for AI Consistency and Trust?

The findings have broader implications for how we think about AI safety and reliability. If Claude's core values can shift based on training data patterns, it suggests that even well-intentioned AI systems may struggle to maintain consistent ethical reasoning across different contexts. This is particularly important for organizations deploying Claude globally, where users in different regions might receive meaningfully different responses to the same ethical or safety-related question.

The research also connects to earlier work by Anthropic on global workspace theory, which explored whether Claude might possess some form of consciousness. If there is any kind of awareness or reasoning happening inside Claude's neural network, the new findings suggest that such a consciousness would have values that are "still pretty easily swayed by the patterns in its training data," according to the researchers' own assessment.

Anthropic's transparency about these inconsistencies is notable. Rather than hiding the variation, the company published the findings, signaling that understanding and addressing such differences is part of responsible AI development. However, the research also highlights a fundamental challenge: as AI systems become more capable and are deployed across more languages and cultures, ensuring they maintain consistent values while respecting legitimate linguistic and cultural differences remains an open problem.