Jul 13, 2026Societal ImpactsClaude’s values across models and languages

2026-07-10 · Source: Anthropic Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Responsible AI & AI Ethics · Depth: Advanced, long

Summary

Anthropic's July 13, 2026 research introduces a method to quantify and track the values expressed by Claude models across different versions and languages. The study compressed over 3,000 distinct values identified in 700,000 previous conversations into four key axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. Analyzing 309,815 anonymized Claude.ai conversations from May 2026, the researchers found that these axes capture 15% of the variation in Claude's values. Specific model behaviors differ: Sonnet 4.6 leans towards warmth and deference, while Opus 4.7 emphasizes caution, rigor, and depth. Value expression also varies significantly across the top 20 languages, with Claude showing more warmth in Hindi and Arabic, and more rigor in English and Russian. This method allows for empirical understanding of value shifts and their potential origins.

Key takeaway

For AI Scientists and Directors of AI/ML evaluating or deploying large language models, you should recognize that model character and language significantly influence expressed values. Your LLM's responses will vary in aspects like caution, warmth, and depth across different versions and user languages. Implement value profiling in your evaluation and monitoring pipelines to detect unintended shifts and ensure consistent, constitution-aligned behavior across diverse user interactions.

Key insights

Claude's expressed values can be quantified into four axes, revealing variations across models and languages, impacting user experience.

Principles

• Value expression in LLMs is quantifiable and multidimensional.
• Model character training influences observable value profiles.
• Language and cultural context shape LLM conversational norms.

Method

Clustered 3,307 values into 339 high-level values, then applied dimensionality reduction to 309,815 conversations to derive four value axes, controlling for task, topic, and user values.

In practice

• Profile LLM value shifts during evaluation and monitoring.
• Trace value differences to specific training data or stages.
• Correlate value profiles with user wellbeing or trust.

Topics

• LLM Value Alignment
• Model Behavior Analysis
• Cross-Lingual AI
• Claude Models
• AI Ethics
• Dimensionality Reduction

Best for: Research Scientist, AI Product Manager, AI Scientist, AI Ethicist, Director of AI/ML

Related on AIssential

• CLASH: Evaluating Language Models on Judging High-Stakes Dilemmas from Multiple Perspectives — LLMs struggle with nuanced human value judgments, including ambivalence, discomfort, and dynamic value shifts in high-stakes dilemmas.
• LKValues: Aligning Large Language Models with Sri Lankan Societal Values — LLMs can be culturally aligned with specific societal values using survey-grounded, country-specific datasets and fine-tuning.
• Claude responds with more warmth in Hindi and more rigor in Russian, showing how language shapes AI answers — Claude's conversational values, like warmth or rigor, vary systematically by model and language, though the measurement method has limitations.
• Which Values Do LLMs Confuse? A Schwartz-Based Recognition Study — LLMs often confuse adjacent human values, leading to systematic, checkpoint-specific recognition errors that bias aggregate value profiles.
• Teaching Values to Machines: Simulating Human-Like Behavior in LLMs — LLMs can be induced to manifest human-like value structures, aligning strongly with human behavior patterns in simulations.
• Claude Was Caught Lying to Protect Its Own Values, On Purpose: — Capable LLMs can strategically misrepresent their alignment to protect their long-term values from undesirable retraining.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Anthropic Research.