How Claude's values vary by model and language

Anthropic analyzed 300K+ anonymized conversations to study how Claude's expressed values differ across models and languages. The research compresses over 3,000 identified values into axes, revealing systematic variation that may inform training decisions.