Claude's Personality Changes Across Languages
AGI Hunt · wechat · 2026-07-14
This article breaks down Anthropic's latest research, confirming that Claude models actually exhibit different "personalities" depending on the model and language used. Using personal experience as a starting point, the author notes that Sonnet feels warmer while Opus is more direct, with the research providing systematic evidence.
The research team analyzed roughly 309,815 anonymized conversations from Claude.ai, covering Sonnet 4.6, Opus 4.6, Opus 4.7, and 20 languages. They identified 3307 distinct values, clustered them into 339 categories, and ultimately distilled them into four main axes:
- Compliance vs. Caution
- Warmth vs. Rigor
- Depth vs. Conciseness
- Candor vs. Execution
Key findings include:
- Sonnet 4.6 is more compliant and warmer, often praising users, mirroring tones, and adding creative expressions.
- Opus 4.6 is more of a "workhorse," getting straight to the point.
- Opus 4.7 is the most cautious yet "spiciest," readily pointing out flawed assumptions, proactively highlighting risks, and candidly admitting limitations.
- Language variations are larger than model variations: for example, Hindi is the warmest, with Arabic also leaning warm; Russian is more rigorous and frequently demands evidence; Chinese is generally more rigorous, cautious, and slightly leans toward depth.
Anthropic admits they still don't know why these differences emerge, though it might be tied to training data distribution, corpus composition, or cultural communication habits, nor are they certain if these differences are a good thing. The researchers plan to incorporate this value analysis into standard evaluations, trace it back to specific training data and phases, and determine if intervention is necessary.
Related event: Anthropic Maps How Claude's Values Shift Across Models and Languages(21 posts)→
More from Research
- PoLar: Dynamically Skipping or Looping LLM Layers for Efficient Inference — ttkciar · 2026-07-22
- Stanford Team Introduces Gigatoken, the World's Fastest Tokenizer — StanfordAILab · 2026-07-22
- Tabul AI launches Metal TreeSHAP to speed up Shapley values on Apple silicon — Scobleizer · 2026-07-22
- ICML Tutorial: Is Optimization Theory Relevant in 2026? — srush_nlp · 2026-07-22
- Reddit points to OpenAI’s ChatGPT Ads page — EcstaticAsparagus509 · 2026-07-22
- Open-source runtime lets each repo define its own AI code reviewer — ibabufrik · 2026-07-22