Large language models today can answer questions in multiple languages and across different cultural contexts. This study explores how well their responses align with the views of people around the world?
The researchers developed an evaluation framework based on data from the international World Values Survey (WVS), which collects public opinions on social, political, and cultural issues across many countries. Using this data, they tested several large language models and compared their responses with how people in different countries actually think about these issues.
The results show that models reflect the views of people from some countries better than others. In certain cases, their responses closely match public opinion in specific countries, while in many others the alignment is significantly weaker.
Interestingly, the language of the prompt has a significant impact on model responses. When a question is asked in a language associated with a particular country or context from the survey, the models tend to align more closely with the views of people from that environment. This suggests that language can serve as a simple way to steer models toward specific cultural perspectives.
The study also finds that models are generally more aligned with contemporary views, while they are less effective at reflecting changes across historical periods.
This research is the first to systematically examine the alignment of AI models with global public opinion across three dimensions - geographical, linguistic, and temporal and it raises an important question: can these models truly represent the diversity of human perspectives, or do they still reflect a limited set of viewpoints?
