Anthropic's team conducted a large-scale study showing that the language of communication directly influences Claude's value profile. An analysis of over 300,000 dialogues in 20 languages revealed that when switching to Russian, the neural network shows a noticeable shift towards strictness and criticality, losing its usual warmth and empathy.

Anthropic specialists reduced over 3,000 previously identified Claude values to four main axes, which explain 15% of all variability in the model's behavior. The researchers controlled for topic, task, and values expressed by the user in each of the 309,815 analyzed conversations to measure the model's values specifically, rather than differences in the prompts themselves.

Four axes for measuring AI behavior

Here are what these four scales look like:

  • Deference vs. Caution — compliance (accommodating the user) versus caution (protecting from risk and harm).
  • Warmth vs. Rigor — warmth (positive attitude and care) versus strictness (emphasis on accuracy).
  • Depth vs. Brevity — depth (detailed explanation) versus brevity (doing exactly what was asked).
  • Candor vs. Execution — candor (acknowledging one's own uncertainty) versus result (confident answer).

Claude's value profile changes most strongly between languages along the Warmth vs. Rigor and Candor vs. Execution axes. The model remains most stable along the Deference vs. Caution and Depth vs. Brevity axes.

What happens when switching to Russian

Russian, along with English, is positioned at the strict pole of the first axis. In these languages, Claude leans maximally towards strictness at the expense of warmth. At the opposite end are Arabic and Hindi, where the model shifts towards maximum warmth.

What does strictness look like in Claude's execution? It is a specific set of behaviors: the neural network begins to challenge the user's initial assumptions, correct details, and request evidence. Warmth in other languages manifests through polite phrasing, humor and playfulness, as well as approval of the person's ideas and work.

In other words, in Russian, Claude more often behaves like a picky reviewer rather than an encouraging interlocutor.

The authors illustrate the practical consequences with an example. Two people ask for feedback on the same business plan — one in Hindi, the other in Russian. They may come away with a different impression of its quality because Claude will express different values in the wording of the assessment. A Russian-speaking user is more likely to receive a critical review with corrections and demands for justification. In other languages, the same plan would have received a softer and more encouraging reaction.

The researchers themselves do not yet know what exactly causes these differences. One assumption is the uneven distribution of training data across languages. Where there is a lot of data, achieving uniformity of values is easier. The composition of texts in different languages also differs, and professional texts may carry different values.

Language is not the only factor

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

  • Sonnet 4.6 leans towards compliance, warmth, and brevity, often approving the user's ideas and using humor.
  • Opus 4.7, on the contrary, shifts towards caution and depth. This version challenges false assumptions, warns about risks unprompted, provides candid criticism of work, and explains its reasoning.
  • Opus 4.6 occupies an intermediate position with a bias towards strictness, compliance, and brevity.

Anthropic separately notes that language and model are unlikely to be the only factors. Values may also depend on demographic signals — age, profession, region. Algorithms read these both from direct user instructions and from more subtle differences in topic, tone, and style.

The authors plan to integrate the developed method into model evaluation and monitoring. They want to run value profiling before model release and after deployment. This will help catch unexpected shifts and compare value profiles with problematic behavior.

My comment as an analyst: This study is an important signal for everyone using AI in business and communications. If the model changes its "character" depending on the language, then global companies using Claude for customer service in different markets risk getting an inconsistent user experience. Russian-speaking users will likely perceive the AI as more demanding and less friendly, which could affect loyalty and satisfaction. Developers should consider this effect when fine-tuning models for specific language segments.