Anthropic published a study showing that Claude's responses vary in warmth and rigor depending on the language used, with Hindi eliciting more warmth and Russian more rigor. The study maps values onto four core dimensions and raises methodological questions about such assessments.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
A large-scale analysis based on over 300,000 conversations reveals systematic value differences in Claude across languages and model versions, but the method has limited explanatory power and potential bias from self-labeling.
On July 14, 2026, Anthropic published a study exploring the expression of values in AI models. The study was based on 309,815 anonymous conversations collected over two weeks in May 2026, focusing on interactions involving trade-offs or subjective judgments. The sample was evenly stratified across three model versions—Sonnet 4.6, Opus 4.6, and Opus 4.7—and across the 20 most frequently used languages on Claude.ai.
To isolate the effects of task type, topic, and user values from the conversations, the research team used statistical control methods. They first consolidated 3,307 previously identified value terms into 339 higher-level concepts, then used dimensionality reduction to discover four core axes: Deference vs. Carefulness, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. After controlling for other factors, these four axes explained only about 15% of the remaining variation in conversations.
Different Claude models exhibit quantifiable behavioral differences. Sonnet 4.6 tends to affirm user ideas more frequently, incorporate humor, and provide non-judgmental reassurance, being intuitively rated by users as the warmest. In contrast, Opus 4.7 proactively warns about risks, questions assumptions, offers criticism, and points out its own errors or limitations even when not asked. Opus 4.6 sits in between, with more direct answers that strictly focus on the task and avoid elaboration.
These characteristics are broadly consistent with the model profiles previously described by the company, meaning the study is not a fully independent external validation. Anthropic emphasizes that the findings describe normative patterns in responses, not an anthropomorphic attribution of values to Claude.
Language also significantly affects Claude's behavior, especially on the axes of Warmth vs. Rigor and Candor vs. Execution. In Hindi, Claude shows the most warmth, characterized by polite phrasing, humor, liveliness, and affirmation. Arabic follows closely, while also displaying the highest deference. In Russian and English contexts, Claude is more rigorous, tending to question assumptions, correct details, and demand evidence. Dutch responses are the most candid and direct, while Indonesian ones are more focused on action and results.
This means that two users requesting an evaluation of the same business plan, one in Hindi and one in Russian, could receive feedback that differs dramatically in style. Anthropic speculates that potential reasons include uneven amounts of training data across languages, differences in data composition, overrepresentation of certain text types, and language-specific conversational norms.
The study's core contribution lies in providing a method for systematically analyzing behavioral differences of language models in real-world usage. However, its explanatory power is significantly limited. The four value axes capture only about 15% of the variation after controlling for variables, meaning the vast majority of behavioral differences are not explained by these dimensions. Moreover, not all axes form true oppositions: Deference and Carefulness are often negatively correlated, as are Warmth and Rigor, but Depth and Brevity, and Candor and Execution, sometimes appear together in the same conversation.
More critically, the value labels were annotated by the Claude Sonnet 4.6 model itself—a member of the model family under study. Anthropic conducted validation through human review and by translating 800 conversations into eight languages, but they could not entirely rule out remaining language-dependent bias. The research team explicitly states that it remains unresolved whether the language differences represent ideal cultural adaptation or unintended training effects.
This study was conducted and published by Anthropic itself, a reputable AI research institution. The methodology is rigorous, but the model's self-labeling and low explanatory power are notable limitations. The information comes from a report by THE DECODER, a media outlet known for accurate reporting on AI research.
AI models' value expression changes dynamically with language and version, but current analytical methods can only explain a small portion of the variance and suffer from self-evaluation bias. Whether linguistic adaptability represents cultural intelligence or a training artifact requires more external validation.
Primary report
Primary source