Claude's tone shifts by language: warm in Hindi, rigorous in Russian

Anthropic published a new study examining how Claude's expressed values shift across different model versions and languages, based on 309,815 anonymized conversations collected over a two-week period in May 2026. The conversations were drawn from instances where Claude had to weigh tradeoffs or make subjective judgments, and the sample was evenly distributed across three Claude models—Sonnet 4.6, Opus 4.6, and Opus 4.7—as well as the 20 most-used languages on Claude.ai.

The research built on Anthropic's earlier study, Values in the Wild, which identified 3,307 individual value terms. The team grouped these into 339 higher-level values and then applied statistical dimensionality reduction to identify patterns in how those values co-occurred across conversations. This analysis yielded four core dimensions: Deference and Caution, Warmth and Rigor, Depth and Brevity, and Candor and Execution. To ensure that measured differences reflected genuine model behavior rather than variation in conversation topics or user-supplied values, Anthropic statistically controlled for task type, subject matter, and user values. The four dimensions account for about 15 percent of the remaining variation after these controls.

The models showed distinct behavioral profiles. Sonnet 4.6 tends to affirm user ideas, leans into humor, and offers comfort without passing judgment. Opus 4.7, by contrast, warns about risks without being asked, questions assumptions, openly critiques, and flags its own mistakes or limits. Opus 4.6 answers more directly, stays close to the task, and avoids extra elaboration. According to Anthropic, these profiles match subjective user impressions—users perceive Sonnet 4.6 as particularly warm, while they more often notice hedging and cautious phrasing from Opus 4.7.

Language differences proved just as striking. Warmth versus Rigor and Candor versus Execution showed the widest variation across languages. Claude expresses the most warmth in Hindi, followed by Arabic, both languages featuring polite phrasing, humor, playfulness, and affirmation. In English and Russian, Claude responds with more rigor, questioning assumptions, correcting details, and asking for evidence. Arabic responses show the most deference overall, while English responses show the most caution. Dutch responses tend to be particularly open and candid, while Indonesian responses lean more toward action and results. Anthropic notes that two people asking Claude to evaluate the same business plan—one in Hindi and one in Russian—could receive feedback that feels very different. The team points to uneven amounts of training data, differences in data composition, overrepresentation of certain text types, and language-specific conversational norms as possible causes.

Anthropologic's analysis presents a systematic method for examining behavioral differences in language models during real-world use, but the method has limits. Not all four axes form true opposites; more deference tended to come with less caution, and more warmth with less rigor, but Depth and Brevity and Candor and Execution could appear together in the same conversation. Additionally, Claude Sonnet 4.6 itself assigned the value labels being studied, meaning a model from the same family whose behavior was being analyzed conducted the labeling. Anthropic verified the method through manual review and by testing 800 conversations translated into eight languages but still does not rule out remaining language-dependent biases. The company explicitly states it is not attributing values to Claude as an agent but rather describing normative patterns in its responses. The results largely align with the model profiles Anthropic has already described, which means this alignment is not an independent check. Whether the language differences represent desirable adaptation to different speech communities or unintended training effects remains an open question.

Anthropic's study builds on earlier research (Values in the Wild) that identified 3,307 value terms, which the company then grouped into 339 higher-level values before using statistical dimensionality reduction to isolate four core axes. By statistically controlling for task type, subject matter, and user-introduced values, Anthropic attempted to separate model behavior from confounding factors—a methodologically sound move that isolates genuine systematic differences. However, the four axes explain only about 15 percent of remaining variation, suggesting that either Claude's behavior is shaped by many other hidden dimensions or that the current framing captures only part of what drives response differences.

The language findings point to a concrete practical issue: training data imbalance and composition differences appear to drive how Claude behaves across languages. Anthropic cites uneven training data, differences in data composition, overrepresentation of certain text types, and language-specific conversational norms as likely causes—all traceable to how the model was built rather than to intentional value alignment. This raises a question about whether these differences are desirable (the model appropriately adapting to different speech communities) or problematic (unintended training artifacts that could mislead users).

A key limitation is that Claude Sonnet 4.6 itself assigned the value labels used to categorize all responses, meaning a model from the same family being studied conducted the analysis. Although Anthropic verified the method through manual review and tested 800 conversations translated into eight languages, this self-measurement approach lacks full independence—the results align closely with Anthropic's own previously published model profiles, which is not an independent check. The company explicitly states it is not attributing values to Claude as an agent but rather describing normative patterns, yet the practical effect for users is indistinguishable from apparent agency.

No comments yet. Be the first to share your thoughts!