A recent study conducted by Anthropic, the creator of Claude, analysed the AI chatbot’s responses in different languages. It found that if you want a friendlier tone, you may want to switch your language settings.
Anthropic compared data from more than 300,000 anonymised conversations in a two-week period during May this year, said The Decoder. The study used the technique “dimensionality reduction” to measure the responses across four separate axes: warmth vs rigour, depth vs brevity, candour vs execution and deference vs caution. Anthropic “explicitly states” that this study does not “imply that Claude intrinsically holds values”, but instead the term “values” refers to “values reflected by Claude’s behaviour and outputs”.
The results showed that Arabic responses had “greater warmth”, meaning they contained more “polite phrasing, humour, playfulness and affirmation”, added The Decoder. On the candour vs execution axis, Dutch responses showed a more “honest appraisal of potential shortcomings”, said The Register.
The study also showed the response differences across Claude’s models. Sonnet 4.6 appeared to be more affirming towards users’ ideas and “offers comfort without passing judgement”, while Opus 4.7 “openly critiques” and “questions assumptions”, according to The Decoder. Sonnet 4.6 is the default free model on the Anthropic platform. Opus 4.7 is the premium paid model of Claude.
These differences remain an “open question”. Anthropic said it was “making sense of this variation” and using what it had learned to “improve Claude’s behaviour”.
|