AI models miss how much everyday norms vary across societies
Researchers compared four AI models with human answers on 150 everyday situations in 90 societies. On average, the models estimated less than half the variation seen between societies in the human study. In some cases they also got which groups held which views wrong. The study does not show that every local-language prompt fails.
What the source reports
A preprint compares four AI models with a human study of everyday norms. The study covers 150 situations in 90 societies. On average, the models estimated less than half the gap between societies that the human answers showed. In some cases, they also assigned views to the wrong societies. The authors caution against replacing local human evidence with model-made cultural summaries. The result does not mean every prompt in a local language will fail in the same way.
Original source
- Title
- Large language models underestimate and partly misrepresent cultural variation in everyday norms
- Author
- Kimmo Eriksson, Irina Vartanova, Pontus Strimling
- Publication
- arXiv
- Date
- Monday, September 28, 2026