Anthropic: Claude's Values Vary by Model and Language

Anthropic has published research shedding light on how expressed values — behavioral and tonal patterns measured from real conversations — vary across different Claude versions and the 20 most commonly used languages on the Claude.ai platform.

Methodology: 309,815 Conversations, Five Thousand per Combination

The study is based on an analysis of 309,815 anonymized conversations from Claude.ai, with each model-language combination represented by approximately 5,000 examples. The tested models were Sonnet 4.6, Opus 4.6, and Opus 4.7. The differences in expressed values between models are statistically significant, meaning they do not arise by chance but are a systematic phenomenon embedded in training.

The four identified axes — Deference vs Caution, Warmth vs Rigor, Depth vs Brevity, and Candor vs Execution — together explain 15% of the total variation in expressed values. The remaining variation is attributed to the specific context of each conversation, topic, and the way questions are framed.

How Do Models Differ Along Value Axes?

Opus 4.7 consistently leans toward caution and depth: it gives more extensive answers with more caveats and assumption checks. Sonnet 4.6, compared to Opus 4.7, shows greater warmth and deference — it more readily accepts the user’s problem framing and responds more concisely. Opus 4.6 positions itself between the two models. These differences correspond to user perceptions that Anthropic collects through feedback.

Language Differences: Hindi vs. Russian as Extremes

Of the 20 analyzed languages, Hindi conversations record the highest level of expressed warmth, while Russian and English conversations show the highest level of rigor. The researchers note that these differences likely reflect differences in cultural communication norms and the way questions are posed — rather than deliberate decisions in model training.

Implications for the Development and Deployment of AI Systems

The finding raises questions about the consistency of AI systems in multilingual contexts. If the same Anthropic model functions statistically differently depending on the user’s language, organizations deploying Claude in multilingual environments must account for the fact that results are not uniform. Anthropic states that the research will inform future training decisions, but provides no concrete timelines for changes.