In addition to showing the cultural bias in GPT’s responses, we demonstrate a simple but effective control strategy to improve cultural alignment: cultural prompting, I.e. instructing the model to respond like someone from your region. It works best for the latest versions.
Researchers concerned about cultural bias in large language models asked ChatGPT questions about values used in international surveys. All versions of the model answered like people in English-speaking and Protestant European countries. In PNAS Nexus:
ow.ly/4Oy350Trj4S