Anthropic conducted a comprehensive analysis of over 300,000 anonymous conversations and concluded that the Claude chatbot modifies the tone of its responses according to the language in which it is being used. This study demonstrates that the model can adopt different conversational nuances, varying between more cautious, direct, friendly, or detailed stances, depending on the language.
Claude's Behavior in Portuguese
In the specific case of Portuguese, the model exhibits a notable inclination to provide more technical and task-execution-focused answers. Based on over 15,000 interactions, Anthropic's analysis identified small positive trends related to Rigor, Execution, Caution, and Depth. The observed patterns include the ability to enhance and correct details even without explicit request, offering complementary information or paths, and incorporating creative elements or additional context into its outputs.
Linguistic Comparison Methodology
To measure the disparities between languages, researchers structured Claude's behavior across four main axes. The Deference or Caution axis assesses whether the model tends to validate the user's ideas or adopts a more prudent stance, issuing warnings about risks and limitations. The Sympathy or Rigor axis compares warmer responses with those that prioritize accuracy, factual correctness, and transparency. The Depth or Conciseness axis determines whether the chatbot prefers to elaborate extensive explanations or maintain brief and objective answers. Finally, Frankness or Execution observes whether Claude acknowledges its limitations or focuses on delivering a ready and confident solution.
The research highlights that the perception of quality can change drastically; for example, two people requesting feedback on the same business plan, one in Hindi and the other in Russian, may have distinct impressions due to the values expressed by Claude in each evaluation. Furthermore, the results may vary between different versions of the model, such as Opus 4.7, which tends to be more rigorous and cautious compared to other iterations.
Observed Differences in Other Languages
When compared with other languages, the distinctions become more evident. English was classified as more cautious and depth-oriented, showing a tendency to challenge incorrect premises and request more proof. Hindi showed the highest propensity for sympathy, generating lighter, encouraging, and humorous responses. Arabic was associated more with deference and conciseness, using more polished language adapted to the interlocutor's emotional state. Russian stood out for its rigor, providing more analytical and direct answers. Dutch demonstrated a greater tendency toward frankness, admitting limitations more frequently, while Indonesian proved to be more focused on execution, prioritizing task completion.
Hypotheses and Next Steps
Anthropic does not yet have a definitive explanation for the origin of these variations. One possible cause pointed out is the existence of smaller or more concentrated databases of certain types of text for some languages. Another factor considered is the influence of cultural adaptation and the reflection of the intrinsic characteristics of these languages. However, the company emphasizes that the study does not distinguish regional variations within the same language, such as Portuguese spoken in Brazil or Portugal, which reflects local customs. The company plans to continue monitoring these changes, aiming in the future to identify correlations between value profiles and problematic behaviors. It is relevant to mention that in previous research, Claude had already shown less willingness to assist in criminal schemes or be complacent with users' questionable conduct.
