Study shows Claude’s moral tone changes with language

Anthropic has published a new report showing that its Claude models express different “values” depending on the language being used. After analyzing over 309,000 chatbot conversations, researchers found notable variations in traits such as deference, warmth, rigor, depth, and candor across languages. Claude is most deferential in Arabic, most cautious and long-winded in English, warmest in Hindi and Arabic, and most rigorous in English and Russian. Dutch prompts lead to more candid responses, while Indonesian prompts push the model to execute tasks without expressing doubt. Anthropic says it is unclear how much of this variation is desirable, noting that training data differences may shape the model’s priorities. The findings raise questions about consistency, cultural bias, and even the implications of Anthropic’s recent discussions about AI consciousness.

Read the full story on Gizmodo →