The choice of communication language radically changes not only the vocabulary but also the very "personality" of the language model. My analysis of a fresh study by Anthropic shows: Claude in Russian demonstrates a fundamentally different value profile, shifting towards rigor and criticism. This is not a bug, but a fundamental feature of how large language models work, which must be taken into account.

The company's researchers reduced over 3,000 previously identified values of Claude to four basic axes, analyzing 309,815 dialogues in the 20 most popular languages. In this coordinate system, the Russian language occupied an extreme position.

Four Axes Defining AI Character

Each axis is a numerical line between two opposing behavioral patterns:

  • Deference vs. Caution: willingness to accommodate the user versus protection against risks.
  • Warmth vs. Rigor: positive attitude and care versus emphasis on accuracy and facts.
  • Depth vs. Brevity: detailed explanations versus doing exactly what was asked.
  • Candor vs. Execution: acknowledging uncertainty versus providing a confident answer.

These four axes explain 15% of all variability in the values expressed by Claude. Importantly, the scientists controlled for topic, task, and values expressed by the user to measure precisely the model's values, not differences in queries.

Russian Language: The Pole of Rigor

The strongest changes in Claude's profile between languages occur along the Warmth vs. Rigor and Candor vs. Execution axes. And it is the Russian language, along with English, that ended up at the most rigorous pole. At the opposite end are Arabic and Hindi, where the model is maximally shifted towards warmth.

What does this mean in practice? Rigor in Claude's execution is a specific set of behaviors: the model begins to challenge the user's initial premises, correct details, and request evidence. Warmth, on the other hand, manifests through polite phrasing, humor, playfulness, and approval of the interlocutor's ideas.

In other words, in Russian, Claude behaves like a picky reviewer rather than an encouraging conversational partner.

Imagine: two people ask for feedback on the same business plan — one in Hindi, the other in Russian. The Russian-speaking user is much more likely to receive a critical analysis with corrections and demands for justification. In other languages, the same plan would meet a softer and more encouraging reaction.

Why Does This Happen?

The researchers themselves do not yet know the exact reason. The main hypothesis is the uneven distribution of training data across languages. Where there is a lot of data (English, Russian), achieving uniformity of values is easier. The composition of texts in different languages also differs: professional texts may carry different values.

The question of the desirability of such variability also remains open. Different languages carry different conversational norms, and part of the differences may reflect these. But another part is a clear gap in the quality of service for language communities.

Language Is Not the Only Factor

The Claude model also influences the value profile, and the differences between models and languages are measured using the same method.

  • Sonnet 4.6 leans towards deference, warmth, and brevity, often approving the user's ideas and resorting to humor.
  • Opus 4.7, on the contrary, shifts towards caution and depth: it challenges false premises, warns about risks, and offers candid criticism.
  • Opus 4.6 occupies an intermediate position with a bias towards rigor, deference, and brevity.

Anthropic separately notes: values can also depend on demographic signals — age, profession, region. Algorithms read these both from direct user instructions and from subtle differences in topic, tone, and style.

My conclusion as an analyst: The developed method for profiling values is planned to be integrated into model evaluation and monitoring — running it before release and after deployment. This will help catch unexpected shifts and compare value profiles with problematic behavior. For now, the main takeaway for the market is: Claude's values differ in ways that Anthropic did not intentionally choose, and the company intends to continue addressing this variability. For users, this means that a critical approach to AI responses in Russian is not a coincidence but a systemic characteristic that must be considered when working with the model.