When Claude answers subjective questions, its replies reflect values defined in its constitution. The paper studies how these values shift across models and languages, aiming to capture “good judgment and sound values that can be applied contextually.”

Researchers analyzed 700,000 anonymized Claude.ai chats, finding more than 3,000 distinct values. To make sense, they compressed them into a few axes—each a line between opposing value groups—so a model’s position shows its leanings.

Applying this method to the top 20 languages and three models (including Opus 4.6, Opus 4.7, and Sonnet 4.6), they identified four key axes that explain 15 % of value variation: Deference–Caution, Warmth–Rigor, Depth–Brevity, and Candor–Execution. For example, Opus 4.6 emphasizes deference and brevity, while Opus 4.7 leans toward caution and depth.

The study demonstrates that Claude’s expressed values differ by model and language, with warmth higher in Arabic and Hindi and rigor in English and Russian. By linking axes to training choices, the work offers a path to evaluate how behavioral training or cultural context shapes AI values.