언어에 따라 달라지는 AI 답변: 클로드의 가치관 분석

Claude responds with more warmth in Hindi and more rigor in Russian, showing how language shapes AI answers Tomislav Bezmalinović Jul 14, 2026 Nano Banana Pro prompted by THE DECODER Key Points Anthropic studied which values Claude models express in real conversations. The company analyzed more than 300,000 anonymized conversations and distilled the observed value patterns into four core axes, including Deference and Caution as well as Warmth and Rigor. The results show systematic differences across models and languages. Sonnet 4.6 responds with more warmth and deference, while Opus 4.7 more often warns about risks unprompted and questions assumptions. The method's explanatory power is limited. The four axes capture only about 15 percent of the variation that remains after statistically controlling for task, topic, and user values. Anthropic also had Claude Sonnet 4.6 assign the value labels, meaning a model from the same family whose behavior was being studied. The company tested for potential language biases but couldn't fully rule out remaining effects. Ask about this article… Search A new Anthropic study maps hundreds of value concepts derived from thousands of individual terms onto four core dimensions. It reveals systematic differences across Claude models and languages, but also raises methodological questions. Anthropic has published a study examining which values Claude expresses in conversations and how those values shift depending on the model and language used. The analysis draws on 309,815 anonymized conversations collected over a two-week period in May 2026. For the value analysis, Anthropic only included conversations where Claude had to weigh tradeoffs or make subjective judgments. The sample was evenly stratified across Sonnet 4.6, Opus 4.6, and Opus 4.7, as well as the 20 most-used languages on Claude.ai. From thousands of value terms to four axes Building on the earlier study Values in the Wild , which identified 3,307 value terms, Anthropic first grouped those into 339 higher-level values. The team then used statistical dimensionality reduction to find patterns in how those values co-occurred. Four core axes emerged: Deference and Caution, Warmth and Rigor, Depth and Brevity, and Candor and Execution. Ad To isolate differences that don't just reflect the conversation topic or user-introduced values, Anthropic statistically controlled for factors like task type, subject matter, and user values. The four axes account for about 15 percent of the remaining variation across conversations after those controls. Ad DEC_D_Incontent-1 Each model has a distinct profile The models differ measurably in how they respond. Sonnet 4.6 tends to affirm user ideas more often, 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 impressions of the models. Users tend to perceive Sonnet 4.6 as particularly warm, while they more often notice hedging and cautious phrasing from Opus 4.7. Ad Language changes the answer The differences across languages are just as striking. Warmth versus Rigor and Candor versus Execution show the widest variation. 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. In Arabic, it shows the most deference. In English, the most caution. Dutch responses tend to be particularly open and candid, while Indonesian responses lean more toward action and results. Two people who ask Claude to evaluate the same business plan, one in Hindi and one in Russian, could receive feedback that feels very different, Anthropic says. The research 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. Ad DEC_D_Incontent-2 Self-measurement with limited explanatory power The study presents an analytical method for systematically examining behavioral differences in language models during real-world use. But its explanatory power has limits. The four axes capture only about 15 percent of the remaining variation. Ad Not all four axes form true opposites, either. More deference tended to come with less caution, and more warmth with less rigor. But Depth and Brevity, along with Candor and Execution, could show up together in the same conversation. There's also the fact that Claude Sonnet 4.6 assigned the value labels, meaning a model from the same family whose behavior was being studied. Anthropic verified the method through manual review and by testing 800 conversations translated into eight languages. The company still doesn't rule out remaining language-dependent biases. Anthropic explicitly states that it isn't attributing values to Claude as an agent but rather describing normative patterns in its responses. The results largely match the model profiles Anthropic itself has described, which means this alignment isn't an independent check. Whether the language differences represent desirable adaptation to different speech communities or unintended training effects remains an open question. AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section. Subscribe now