Anthropic study finds Claude AI replies differ by language
Technology
Anthropic's new study shows Claude AI responds differently based on language.
After looking at more than 309,000 conversations across three models (Sonnet 4.6, Opus 4.6, and Opus 4.7), researchers found that Claude replies in Hindi and Arabic were much warmer (think humor and friendly reassurances), while answers in English and Russian felt more analytical and precise.
Sonnet 4.6 friendliest, Opus 4.7 serious
Among the models tested, Sonnet 4.6 was the friendliest, often offering reassurance and jokes. Opus 4.7 leaned into logic and risk assessment, making it sound more serious.
Anthropic says the results reflect differences in how the AI responds across languages and models, not different beliefs, and could help improve future AI consistency.