New research from AI company Anthropic reveals that its AI assistant, Claude, exhibits distinct behavioral patterns and expresses different values depending on the specific model version and the language used in conversations. The findings are based on an analysis of over 309,000 anonymized user interactions.
Researchers distilled more than 3,300 identified values into four key behavioral dimensions: deference versus caution, warmth versus rigor, depth versus brevity, and candor versus execution. The study controlled for conversation tasks, topics, and user-expressed values to isolate Claude's inherent responses.
According to Anthropic, different Claude models displayed unique profiles. For instance, Sonnet 4.6 leaned towards warmth, deference, and brevity, often using humor. In contrast, Opus 4.7 emphasized rigor, caution, candor, and depth, frequently challenging assumptions and acknowledging limitations. Opus 4.6 generally adopted a more concise, execution-focused approach with an emphasis on rigor.
Language also played a significant role in Claude's responses. Arabic conversations tended to be more deferential and warmer, while English conversations prioritized caution and rigor, often involving detailed explanations and challenges to assumptions. Russian responses were also rigorous, while Dutch responses were noted for their candor in acknowledging uncertainty. Indonesian responses focused on task completion.
Anthropic emphasized that this research does not suggest Claude possesses values itself. The company is still investigating the causes of these behavioral differences and whether they are desirable. This framework, however, could be useful for evaluating future models and detecting unintended behavioral shifts. Previous Anthropic studies have explored Claude's "functional introspective awareness" and internal "emotion vectors."
