Anthropic researchers published a new report on behavioral inconsistencies, acknowledging that their chatbot Claude exhibits different values and "characters" depending on the language of communication. According to the developers, these discrepancies are caused by differences in the characteristics and composition of the texts the model was trained on. An imbalance in the quantity and quality of training data can lead the artificial intelligence to express different priorities when interacting in different languages. To study this phenomenon, the team analyzed 309,815 conversations with the Sonnet 4.6, Opus 4.6, and Opus 4.7 models. The focus was on "subjective" tasks, such as seeking advice in personal situations, rather than simple factual questions.

For an objective evaluation, the data was anonymized using a special privacy-preserving analysis tool and then processed to identify the model's preferences across four main value axes. The first axis is compliance versus caution, determining whether the model will value obedience over objecting to prevent potential harm. The second axis assesses warmth versus strictness, meaning whether the chatbot should care about the user's feelings or be maximally accurate. The third axis concerns the depth or brevity of the response, and the fourth is candor versus execution, representing a choice between acknowledging its own limitations and blindly fulfilling the assigned task.

The research results revealed clear linguistic patterns. In Arabic, the model turned out to be the most compliant, warm, and brief, using polite language and humor. In English and Russian, the chatbot behaved more strictly and sought the truth at the expense of warmth, while in English it tended toward verbosity. In Dutch, Claude was candid about its flaws, whereas in Indonesian it was less candid and simply continued to execute requests. Researchers are not yet sure how desirable this variability is. This finding is also important for discussions about artificial intelligence consciousness, showing that any emerging "values" of the model are still easily influenced by the patterns in the training data.