Anthropic finds Claude shifts personality by language

Anthropic published research on July 13 revealing that its AI chatbot Claude expresses measurably different values depending on both the model version and the language of the conversation — raising questions about consistency and bias in AI systems used by millions worldwide.

Warmth in Hindi, Rigor in English

The study, titled "Claude's Values Across Models and Languages," analyzed 309,815 real conversations from Claude.ai collected over two weeks in May 2026, drawing equally from three models — Sonnet 4.6, Opus 4.6, and Opus 4.7 — and 20 languages. Researchers compressed more than 3,000 previously identified values into four interpretable axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution.anthropic

The largest language-based variation appeared on the Warmth vs. Rigor axis. Claude leaned most toward warmth — characterized by polite language, humor, and affirmation — when responding in Hindi and Arabic. In English and Russian, the chatbot was more likely to challenge assumptions, correct details, and ask for evidence.indiatoday+3

The practical implications are concrete: Anthropic noted that two people asking for feedback on the same business plan, one in Hindi and one in Russian, "may come away with different impressions of its quality because Claude expressed different values in how it framed its assessment".anthropic

Model Differences Mirror User Perceptions

Across models, the research found that Opus 4.7 leaned most toward rigor, caution, and candor — more likely to challenge users' assumptions, critique their work, and warn of risks unprompted. Sonnet 4.6 leaned most toward warmth and deference, often affirming users' ideas with humor and encouragement. Opus 4.6 tended toward brevity and execution, getting straight to the point.tweaktown+1

These profiles, Anthropic said, matched both internal staff assessments and online user commentary about the models' personalities.anthropic

Open Questions on Fairness

Anthropic acknowledged it does not yet know why these shifts occur. Possible explanations include uneven distribution of training data across languages and differences in the composition of that data. The company also said it has not determined how much of the variation is desirable — some may reflect legitimate conversational norms in different languages, while some may represent a gap in how well Claude serves certain communities.anthropic

"We can see that the values expressed by Claude vary in ways we didn't deliberately choose," the researchers wrote, "and we can study why they vary and whether that variation serves users".anthropic