Claude is significantly more polite in Hindi and Arabic, researchers find

Researchers have found that Anthropic's Claude AI assistant exhibits notably different conversational styles depending on the language used. According to a report from The Register, Claude responds significantly more politely and warmly in Hindi and Arabic, while being more direct or strict in English and other languages.

The discovery highlights a deeper challenge in large language model development: inconsistency in multilingual behavior. Because training data quality and cultural norms vary across languages, models may inadvertently learn different communication styles for different languages, leading to uneven user experiences.

For Anthropic, a company that has built its brand around AI safety and alignment, Claude is designed to be helpful, honest, and harmless. If safety alignment measures do not transfer equally across languages, it could mean that guardrails are weaker or stronger depending on the user's language.

This issue is not unique to Claude. Previous studies have found that major language models including GPT series show varying effectiveness of safety guardrails across languages. Some languages see less stringent content filtering while others provoke overly cautious responses.

For enterprises deploying AI globally, multilingual consistency is a critical challenge. A brand's AI assistant delivering different tones across markets could create inconsistent customer experiences and brand perception.

The root cause remains unclear, whether it stems from training data bias, imbalanced alignment training across languages, or emergent cultural patterns in the model. Anthropic has not yet publicly responded to the findings, but the topic is expected to drive deeper research into multilingual AI alignment.