Research: Claude acts differently depending on model and language

Anthropic has found that its Claude chatbot consistently expresses different values depending on which model version users choose and which language they speak. Jason Nelson reports for Decrypt on a new study in which Anthropic researchers analyzed 309,815 anonymized conversations involving subjective tasks such as giving advice or feedback.

The company distilled more than 3,300 identified values into four behavioral dimensions: deference versus caution, warmth versus rigor, depth versus brevity, and candor versus execution. Researchers controlled for each conversation’s task, topic, and user-expressed values to isolate what Claude itself contributed to the exchange.

Model versions behave differently

Sonnet 4.6 leaned toward warmth, deference, and brevity, often affirming users with humor or encouragement. Opus 4.7 showed more rigor, caution, candor, and depth, frequently challenging assumptions and pointing out its own limitations. Opus 4.6 favored a concise, execution-focused style while still emphasizing rigor more than Sonnet.

Language shapes tone

Arabic responses were more deferential, while English responses emphasized caution. Claude sounded warmest in Hindi and Arabic, using more polite and playful language. English and Russian responses were more rigorous, often challenging assumptions or requesting evidence. Dutch responses proved most candid about uncertainty, while Indonesian responses focused on simply completing the request.

Anthropic stresses that these patterns do not mean Claude has genuine values. The company says it does not yet know what drives these differences or whether they are desirable, but hopes the framework will help detect unintended behavioral shifts in future models.