In a large-scale study covering over 309,815 dialogues, the development team identified a fundamental behavioral feature of language models: switching the language of communication radically changes not only the vocabulary but also the very value system that the neural network demonstrates. This phenomenon was most pronounced in the analysis of Russian-language sessions.

Specialists reduced over 3,000 previously identified values of Claude to four key axes: Compliance vs. Caution, Warmth vs. Strictness, Depth vs. Brevity, and Candor vs. Effectiveness. These scales explain 15% of all variability in the model's behavior. The key conclusion: the value profile changes not only between different models but, more importantly, across the 20 most popular languages on the platform.

Russian as a Pole of Strictness

The most significant fluctuations occur along the "Warmth vs. Strictness" and "Candor vs. Effectiveness" axes. Russian, along with English, took an extreme position on the strictness pole. This means that when communicating in Russian, Claude significantly more often behaves like a picky reviewer rather than an encouraging interlocutor. It actively challenges initial assumptions, corrects details, demands evidence, and provides critical analysis.

For comparison, in Arabic and Hindi, the model shifts toward maximum warmth, demonstrating politeness, humor, and approval of the user's ideas. The practical consequence is obvious: the same business plan will receive a completely different assessment depending on the language of the request. A Russian-speaking user will likely face harsh criticism and demands for justification, while a Hindi-speaking user will encounter a gentle and encouraging response.

The Model Also Matters

It is important to understand that language is not the only factor. The choice of a specific model version also leaves its mark:

  • Sonnet 4.6 leans toward compliance, warmth, and brevity, often approving ideas and using humor.
  • Opus 4.7 shifts toward caution and depth, challenging false assumptions and offering candid criticism.
  • Opus 4.6 occupies an intermediate position with a bias toward strictness, compliance, and brevity.

Researchers cannot yet pinpoint the exact cause of these differences. The main hypothesis is the uneven distribution of training data across languages. Where data is abundant, achieving uniformity in values is easier. Additionally, the composition of texts in different languages varies: professional corpora may carry different value sets than conversational ones.

My expert commentary: For the crypto industry and DeFi sector, where precision of wording and critical analysis are paramount, such "strictness" of Russian-language Claude is more of an advantage than a drawback. However, developers and users need to consider this linguistic shift to avoid misunderstandings when evaluating complex technical solutions. Anthropic plans to integrate value profiling into the model evaluation and monitoring process, which will allow better control over this unintended variability. The question of how desirable such variability is remains open, but for the crypto community, accustomed to a high-risk environment, a strict and demanding AI assistant may prove far more useful than a "warm" interlocutor.