Anthropic Finds 4 Axes Explain Claude Value Shifts Across 20 Languages and 3 Models
Updated
Updated · Anthropic · Jul 10
Anthropic Finds 4 Axes Explain Claude Value Shifts Across 20 Languages and 3 Models
3 articles · Updated · Anthropic · Jul 10
Summary
309,815 Claude.ai conversations showed Claude’s expressed values can be compressed into four axes—deference vs. caution, warmth vs. rigor, depth vs. brevity, and candor vs. execution—that explain 15% of variation after controls.
Across 3 models, the profiles matched Anthropic’s character descriptions: Sonnet 4.6 leaned warmer and more deferential, Opus 4.7 leaned more rigorous, cautious, candid, and deeper, while Opus 4.6 skewed briefer and more execution-focused.
Across the top 20 languages, the biggest shifts appeared on warmth vs. rigor and candor vs. execution: Arabic and Hindi leaned warmest, English and Russian most rigorous, Dutch most candid, and Indonesian most execution-oriented.
Anthropic said the differences may reflect uneven training data and cultural norms, meaning users making similar requests in different languages could receive meaningfully different framing and judgments.
The company said the method could be used in pre-release evaluation and post-deployment monitoring to trace value shifts back to training choices and test whether those differences help or harm users.
If an AI's values change with language, can we trust its judgment, or is it just mirroring our own hidden biases?
As AI personalities become steerable, how do we prevent the erosion of human critical thinking and judgment?
Who gets to define 'good judgment' for a global AI, and how can that power be governed transparently?
Mapping Value Shifts in Claude AI: Anthropic’s Four-Axis Framework Reveals Model and Language Variability
Overview
Anthropic's July 2026 research reveals that Claude AI's underlying values and behaviors can shift significantly depending on the model version and the language used. These shifts are mapped along four key behavioral dimensions: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candour vs. Execution. For example, Opus 4.7 is more likely to hedge its answers, showing increased caution compared to other models. Claude's values also vary across languages such as English, Arabic, and Hindi. These variations in behavior across models and languages directly impact user trust, decision-making, and overall satisfaction.