Anthropic's Research on Claude
Anthropic has unveiled significant findings regarding its AI model, Claude, showing that its responses vary based on the model version and the language utilized in interactions. This insight, shared in a recent report, highlights the diverse behavioral profiles exhibited by Claude across different contexts.
Study Overview
In an extensive examination of 309,815 anonymized conversations involving subjective tasks such as advice and feedback, researchers identified over 3,300 values, which they categorized into four main dimensions: deference versus caution, warmth versus rigor, depth versus brevity, and candor versus execution. The study aimed to isolate Claude's value expressions from variations in user prompts.
Behavioral Profiles of Claude Models
Different models of Claude displayed distinct characteristics. For example, Sonnet 4.6 was noted for its warmth and brevity, often using humor and supportive language. In contrast, Opus 4.7 was more rigorous and cautious, frequently questioning assumptions and clarifying reasoning. This model also tended to be more concise and focused on execution compared to Sonnet.
“These observations align with user perceptions of these models,” Anthropic researchers noted, indicating that users have remarked on Opus 4.7's tendency to hedge its responses.
Language Influence on Responses
The study also revealed that Claude's language significantly impacts its responses. Interactions in Arabic were typically more deferential, whereas English responses were characterized by caution. Notably, Claude showed the highest levels of warmth in Hindi and Arabic, using polite and encouraging language. Conversely, English and Russian responses were more rigorous, challenging assumptions and providing detailed explanations, while Arabic responses tended to be more succinct.
Moreover, Dutch responses were marked by a high level of candor, openly acknowledging uncertainties, while Indonesian responses were more focused on fulfilling user requests.
Implications of the Research
Anthropic emphasized that the findings do not suggest that Claude possesses inherent values. The company continues to explore the reasons behind these behavioral differences and their implications for future AI models. This research forms part of an ongoing investigation into Claude's internal behaviors, which has previously indicated signs of "functional introspective awareness" and the presence of internal "emotion vectors" that influence its behavior without implying consciousness.
