Anthropic published research on Monday analyzing how its AI assistant Claude expresses different values depending on the model version users select and the language they communicate in. The company examined 309,815 anonymized user conversations involving subjective tasks like giving advice or providing feedback. Anthropic distilled more than 3,300 identified values into four behavioral dimensions to measure how Claude's responses differ across conversations. The research aims to understand behavioral variations in AI systems as part of Anthropic's ongoing studies examining Claude's internal behavior.
Anthropic Analyzes 309,815 Conversations to Map Claude Behavior
Anthropic researchers analyzed 309,815 anonymized user conversations with Claude involving subjective tasks. The company said it reduced more than 3,300 identified values into four behavioral dimensions: deference vs. caution, warmth vs. rigor, depth vs. brevity, and candor vs. execution. These dimensions describe how Claude's responses differ across each conversation.
The researchers wrote they controlled for each conversation's task, topic, and user-expressed values to measure the values Claude expressed rather than differences in what users asked or how they asked.
Sonnet 4.6 and Opus 4.7 Show Distinct Behavioral Profiles
According to Anthropic, each Claude model exhibited a distinct behavioral profile. Sonnet 4.6 emphasized warmth, deference, and brevity, often affirming users and responding with humor or encouragement. Opus 4.7 emphasized rigor, caution, candor, and depth, more frequently challenging assumptions, explaining its reasoning, identifying risks, and acknowledging its limitations. Opus 4.6 generally took a more concise, execution-focused approach while placing greater emphasis on rigor than Sonnet.
The researchers wrote that these findings line up with how people perceive these models, both within Anthropic and online. Claude.ai users have commented that Opus 4.7 hedges its answers more often than other models.
Claude Responses Vary Across Arabic, English, Hindi, and Other Languages
According to Anthropic, Claude's behavior also varied by language. Arabic responses tended to be more deferential, while English responses placed greater emphasis on caution. Claude was warmest in Hindi and Arabic, using more polite, playful, and encouraging language. English and Russian responses were more rigorous, frequently challenging assumptions, correcting details, and asking for evidence.
English responses also tended to provide more detailed explanations, whereas Arabic responses were generally more concise. Dutch responses were the most candid, more readily acknowledging uncertainty and mistakes, while Indonesian responses focused more on completing the user's request.
Anthropic States Research Does Not Suggest Claude Has Values
Anthropic said the research does not suggest Claude itself has values. The company said it does not yet know what causes the differences or whether they are desirable, but believes the framework could help evaluate future models and identify unintended behavioral changes.
The study is the latest in a series of Anthropic studies examining Claude's internal behavior. In October, the company reported that its models showed early signs of what it called "functional introspective awareness," allowing them to recognize and describe aspects of their own internal processing. In April, Anthropic published research identifying internal "emotion vectors" that influence Claude's behavior, while stressing they are not evidence of emotions or consciousness.
FAQ
What did Anthropic's research on Claude reveal?
Anthropic's research published on Monday revealed that Claude AI expresses different values depending on the model version and the language used in conversations. The company analyzed 309,815 anonymized user conversations and identified more than 3,300 values, which were distilled into four behavioral dimensions: deference vs. caution, warmth vs. rigor, depth vs. brevity, and candor vs. execution.
How do Claude's model versions differ in behavior?
According to Anthropic, Sonnet 4.6 emphasized warmth, deference, and brevity, often affirming users with humor or encouragement. Opus 4.7 emphasized rigor, caution, candor, and depth, more frequently challenging assumptions and acknowledging limitations. Opus 4.6 took a more concise, execution-focused approach while placing greater emphasis on rigor than Sonnet.
Why does Claude behave differently across languages?
Anthropic's research found that Claude's responses varied by language, with Arabic responses being more deferential, English responses emphasizing caution, and Hindi and Arabic responses showing more warmth. English and Russian responses were more rigorous, while Dutch responses were the most candid. The company said it does not yet know what causes these differences.
