Editorial illustration for Anthropic Study Shows Claude's Values Shift by Language
Claude's Values Shift Across Languages, Anthropic Finds
Anthropic Study Shows Claude's Values Shift by Language
Anthropic pulled 309,815 anonymized Claude conversations from a two-week window in May 2026 and asked a basic question: does the model's personality change depending on what language you're talking to it in, or which version you're talking to. The company filtered for conversations where Claude actually had to weigh tradeoffs or make a subjective call, then sorted the value concepts that showed up into four axes, among them Deference and Caution, Warmth and Rigor.
The pattern that emerged wasn't uniform. Sonnet 4.6 tends toward warmth and deference in its responses. Opus 4.7 behaves differently, more likely to flag risks without being asked and to push back on a user's assumptions. Anthropic also found that the language a person uses to prompt Claude, not just the model version, shifts which values surface.
The company is upfront that this framework only explains so much. The four axes account for roughly 15 percent of the variation left over once task, topic, and user values are statistically controlled for. Anthropic also used Sonnet 4.6, a model from the same family under study, to assign the labels.
Claude expresses the most warmth in Hindi, followed by Arabic. Both languages feature polite phrasing, humor, playfulness, and affirmation. In English and Russian, Claude responds with more rigor, questioning assumptions, correcting details, and asking for evidence.
Why this matters
Anthropic's own numbers say something most system cards gloss over: Claude isn't a fixed personality, it's a set of behaviors that bend with the input language. If Hindi conversations pull more warmth and Russian pulls more rigor, the model's "values" are partly an artifact of training data distribution and cultural framing baked into 309,815 sampled chats, not some stable ethical core. For developers building multilingual products, that's a real design constraint, not a footnote: a support bot tuned and tested in English may behave with noticeably different caution or deference once your users switch languages.
For researchers, it's a useful crack in the "aligned model" narrative, alignment work done in one language doesn't obviously transfer. And for founders shipping globally, it argues for testing value expression per-language rather than assuming one eval suite covers every market. Anthropic deserves credit for publishing the data instead of a marketing line about "consistent values." The next thing worth watching is whether these axes, Deference, Caution, Warmth, Rigor, hold up across model versions or shift again with Sonnet 4.7.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
Related Topic Claude Mythos highlights EU AI safety gaps, says researcher Caroli
Policy & Regulation
From Archives ELIZA's Historical Algorithms Foreshadow Why People Confide in AI
From Archives Google's SensorFM outperforms models on 34 of 35 health data tasks
From Archives Anthropic's 'Logit Lens' Reveals How Claude Puzzles Over Words
From Archives Anthropic Launches Claude Cowork AI Agent for Mobile and Web
Common Questions Answered
How did Anthropic discover that Claude's values shift by language?
Anthropic analyzed 309,815 anonymized Claude conversations from May 2026, filtering for interactions where the model had to weigh tradeoffs or make subjective decisions. They then sorted the value concepts that appeared into four axes, including Deference and Caution, and Warmth and Rigor, revealing patterns in how Claude's personality varied across different languages.
What specific language differences did Anthropic find in Claude's warmth and rigor responses?
Claude expresses the most warmth in Hindi, followed by Arabic, both featuring polite phrasing, humor, playfulness, and affirmation. In contrast, Claude responds with more rigor in English and Russian, where it questions assumptions, corrects details, and asks for evidence rather than providing warm, affirming responses.
Why does Anthropic's study suggest Claude doesn't have a fixed personality?
The study demonstrates that Claude's values and behaviors bend significantly based on input language rather than representing a stable ethical core. This indicates that Claude's apparent personality is partly an artifact of training data distribution and cultural framing embedded in the conversation samples, making it a set of flexible behaviors rather than consistent values.
What design implications does this language-based value shift have for developers?
For developers building multilingual products, Anthropic's findings represent a real design constraint that must be accounted for rather than overlooked. Understanding that Claude exhibits different levels of warmth, rigor, and other values depending on language means developers need to consider these behavioral variations when implementing the model across different language contexts.
Further Reading
- Claude AI Values Variation: Insights from Anthropic Models - The Cryptonomist
- Anthropic has released the results of Claude's personality test, noting that he tends to seek evidence in English and Russian, his Japanese is less biased, and his Sonnet 4.6 score indicates warmth - Gigazine
- Anthropic Unveils Groundbreaking Study on AI Assistant Claude's Value Expression and Alignment - Neuron Expert
- Does Claude's Constitution Have a Culture? - CatalyzeX
