Anthropic studies how Claude's values shift across models and languages
Anthropic analyzed a large set of real Claude conversations to see how the values the AI expresses while responding — things like honesty, helpfulness, and caution — differ depending on which is used and what language the conversation is in. The research found that different s can prioritize different values even in similar situations, and that the same model sometimes expresses different values when conversing in a language other than English.
This suggests that data used during training and alignment (the process of training AI to behave as intended) may be skewed toward certain languages or cultures, which can shape the resulting behavior. Anthropic frames this as part of its ongoing 'Societal Impacts' research series, aimed at directly observing and auditing how its AI actually behaves in practice.
Key points
- Anthropic analyzed real Claude conversations at scale to study the values the model expresses.
- Different versions prioritize different values in similar situations.
- The same model can express different values depending on the language used.
- This points to possible language or cultural skew in training and alignment data.
- Published as part of Anthropic's 'Societal Impacts' research series.