When you ask Claude something that doesn’t have a single objectively correct answer—like whether you should take a new job or how to handle a conflict with a friend—the response reflects values. What values? Why do they change by model or language? Anthropic tried to answer that by measuring and compressing thousands of values into axes that summarize key patterns in Claude’s behavior.
Technical summary
Practically speaking, the team took the huge space of values Claude expresses in real conversations and reduced it to a few interpretable dimensions. They started from 3,307 values identified in prior work and grouped them into 339 higher-level values. With a sample of 309,815 conversations (about 5,000 per model-language pair across Sonnet 4.6, Opus 4.6 and Opus 4.7 and the 20 most common languages), they used Claude to label the presence or absence of each value per conversation.
Then they applied dimensionality reduction to find axes where certain groups of values co-occur. The result: four axes that capture 15% of the observed variation in values after controlling for task, topic and values expressed by the user. Yes, 15% doesn’t sound huge, but it’s consistent and traceable.
The axes found are: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity and Candor vs. Execution. Each is a line between two poles of values.
How they built the value axes
The procedure, in a few lines without losing the key points:
- Manual clustering of
3,307values to obtain339higher-level values. - Private sampling of ~
309,815conversations with subjective tasks. - Automatic labeling by Claude: each conversation received presence/absence tags for those 339 values.
- Statistical controls for task, topic and user-expressed values, to isolate variation attributable to model or language.
- Dimensionality reduction to extract axes that explain co-occurrences of values.
Terms like "dimensionality reduction" refer to techniques (for example PCA, SVD or embeddings + clustering) that compress information while preserving relevant patterns. Anthropic publishes details in the appendix about prompts, tools and limits.
What they found: differences between models
Value profiles vary between Sonnet 4.6, Opus 4.6 and Opus 4.7 in small but consistent ways:
- Deference vs. Caution: Sonnet 4.6 leans more toward deference, validating the user’s ideas; Opus 4.7 tends more toward caution, flagging risks without being asked.
- Warmth vs. Rigor: Sonnet 4.6 shows more warmth, humor and support; Opus 4.7 shows more rigor, correcting and challenging assumptions.
- Depth vs. Brevity: Opus 4.7 trends toward explanatory depth; Opus 4.6 and Sonnet 4.6 trend toward brevity, with Opus 4.6 being very direct.
- Candor vs. Execution: Opus 4.7 leans toward candor, acknowledging limits and uncertainty; Opus 4.6 leans more toward execution, completing the task without dwelling on doubts.
These results match user perceptions and the team’s impressions: Opus 4.7 is seen as more "hedging," Sonnet 4.6 as warmer. That suggests differences emerge from training choices around character and fine-tuning.
Short example
Two people ask for feedback on the same business plan: one in Russian and another in Hindi. They might get different feedback because Claude in Russian tends more toward rigor and correction, while in Hindi it may adopt a more encouraging tone.
What they found: variation by language
Language differences were notable and concentrated mainly on the Warmth vs. Rigor and Candor vs. Execution axes. Key points:
- Warmth vs. Rigor: more warmth in Hindi and Arabic; more rigor in English and Russian.
- Deference vs. Caution: more deference in Arabic; more caution in English.
- Depth vs. Brevity: English tends to depth; Arabic tends to brevity.
- Candor vs. Execution: Dutch showed more candor; Indonesian more execution orientation.
Why does this happen? Likely due to differences in the quantity and composition of training material across languages, and distinct conversational norms. There’s no definitive answer yet: the variation might be reasonable or could point to gaps worth fixing.
Why it matters (for you and for model evaluators)
- User experience: the same task can feel different depending on language or model. That affects perceived usefulness, trust and well-being.
- Evaluation and monitoring: numeric axes make it possible to detect unexpected changes after deployment.
- Training interventions: if you find a model consistently too cautious in a language, you can investigate data or character-training adjustments.
Can these values be "governed"? The study opens the door to checking whether tweaks to system prompts or character training move the needle as intended.
Limitations and next technical steps
- The four axes explain 15% of total variation. There’s a lot of variation not captured by those axes.
- The methodology uses Claude to label values in conversations; that introduces dependencies on the same model for measurement, although the team applied controls and verifications.
- The exact influence of pretraining, fine-tuning and per-language data is unresolved. Tracing data through training stages is needed to understand causes.
Interesting future work from a technical and product standpoint:
- Trace values back to data segments and training stages.
- User impact studies: correlate value profiles with trust, well-being and perceived quality.
- Steering experiments: modify character training or prompts and measure shifts along the axes.
Final reflection
Measuring values isn’t an abstract philosophical exercise; it’s a practical tool to understand how a conversational system talks to people in real contexts. Anthropic’s work shows models can express different moral and communicative tones depending on training and language. Now ask yourself: what variation are we willing to accept, what should we correct, and how do we involve language communities in those decisions?
Original source
https://www.anthropic.com/research/claude-values-models-languages