An Anthropic study has found that its AI assistant Claude exhibits different behavioural traits depending on the language users interact with, suggesting that AI responses are influenced more than just by translation.
The research analysed over 3.7 million Claude conversations across 45 languages, making it one of the largest studies to examine how large language models behave in multilingual settings. Researchers manually reviewed thousands of conversations before training AI classifiers to identify behavioural patterns.
According to the study, Claude tends to adopt a warmer, more empathetic tone in Hindi, placing greater emphasis on encouragement and emotional support. In contrast, responses in English were far more rigorous and analytical, with the AI focusing more on accuracy, precision and structured reasoning.
The researchers also observed behavioural differences across other languages. Arabic responses leaned more towards directness, while Japanese conversations displayed greater caution and humility. These variations remained consistent even when users asked similar questions in different languages.
Anthropic says the findings suggest language influences the behavioural patterns AI models develop during training, rather than simply changing how responses are translated. The company believes understanding these differences will help improve AI fairness, consistency and multilingual performance as language models are increasingly deployed worldwide.
The study also introduces a new benchmark called “Anthropic 1M”, which researchers say could help developers better evaluate multilingual AI systems and ensure more reliable behaviour across languages.
