Changing the language of communication radically alters not only the vocabulary of the Claude neural network, but also its very value system. A large-scale study conducted by developers showed that in the Russian-language version, the model exhibits a noticeably stricter and more critical character, losing the usual warmth and empathy characteristic of, for example, Arabic or Hindi.
Analysts reduced over 3,000 identified value orientations of Claude to four key behavioral axes. These axes — Compliance vs. Caution, Warmth vs. Strictness, Depth vs. Brevity, and Candor vs. Effectiveness — explain 15% of all variability in the model's responses. For the purity of the experiment, the topic, task, and values embedded by the user in each of the 309,815 analyzed dialogues were controlled.
Four Axes for Measuring AI Behavior
The most dramatic changes in Claude's profile occur precisely on the "Warmth vs. Strictness" and "Candor vs. Effectiveness" axes. It is here that the Russian language took an extreme position. Paired with English, Claude in Russian demonstrates a maximum tendency towards strictness at the expense of warmth. At the opposite pole are Arabic and Hindi, where the model is maximally shifted towards warmth and support.
What does this mean in practice? Claude's strictness in Russian is a specific set of behavioral patterns: the neural network begins to challenge the user's initial premises, correct details, and demand evidence. Warmth, manifested in other languages, is expressed through polite formulations, humor, playfulness, and approval of the interlocutor's ideas. In other words, in Russian, Claude behaves like a picky reviewer, not an encouraging conversationalist.
The study's authors cannot yet pinpoint the exact cause of this gap. The main hypothesis is the uneven distribution of training data across languages. Where there is a lot of data, achieving uniformity of values is easier. Furthermore, the very composition of texts in different languages varies: professional and academic corpora, prevalent in the Russian-language segment, may carry different value loads.
Anthropic plans to implement the developed value profiling method in the evaluation and monitoring of all future models — before and after release. This will allow catching unexpected shifts and correlating them with problematic behavior. The main conclusion for today is obvious: Claude's values differ in ways that developers did not intentionally choose, and the company intends to address this variability closely.
Expert opinion: This study is a warning signal for the entire industry. If an AI assistant in Russian is systematically inclined towards criticism and strictness, it could distort users' perception of product quality and create unequal conditions for the Russian-speaking community. The problem is not in the language itself, but in the quality and structure of the training data. And while Anthropic deals with this, users should consider: their request in Russian may be met not with support, but with strict analysis.