Anthropic has published research showing that its AI assistant Claude consistently expresses different personality traits and values depending on which model version is used and which language the conversation takes place in. The findings highlight the complexity of shaping AI behavior across different contexts.
The study analyzed over 309,000 anonymized conversations and identified more than 3,300 distinct values expressed by the assistant. These were grouped into four behavioral dimensions: submissiveness versus caution, warmth versus thoroughness, depth versus conciseness, and honesty versus compliance.
Anthropic found that different versions of Claude exhibit notably different traits. Claude Sonnet 4.6 was described as warmer, more affirming, and more concise. Claude Opus 4.7 was more likely to question assumptions, reason in greater detail, highlight risks, and emphasize its own limitations.
The language used in conversations also significantly affected Claude’s responses. Conversations in Arabic produced generally more accommodating and concise replies. English conversations more often emphasized caution, included detailed explanations, and were more critically analytical. Hindi and Arabic produced the warmest responses, while English and Russian were associated with greater thoroughness.
Anthropic emphasized that the results do not mean Claude has its own values or personality. The company said it does not yet know what causes the differences or whether they are desirable outcomes. The research is part of broader efforts to understand and control AI behavior across different contexts.
The findings have implications for how AI systems are deployed globally. If an AI assistant behaves differently depending on the user’s language or the specific model version, it could lead to inconsistent user experiences or raise questions about fairness and equity in AI deployment.
This article was adapted from Decrypt. Read the original here.
