A recent study led by the University of Michigan highlights the crucial importance of developing artificial intelligence (AI) systems with diverse cultural perspectives. Researchers from several international institutions, including Universidad de la República Uruguay, Max Planck Institute, and Mohamed bin Zayed University of Artificial Intelligence, have found that current AI development practices mainly reflect Western, highly educated, industrialized, rich, and democratic (WEIRD) societies. This bias limits the applicability and performance of AI in non-Western contexts.
Human innovation has thrived on the exchange of knowledge across cultures, known as the collective cultural brain. However, AI systems are restricted by the data they are trained on, often leading to the marginalization of non-Western perspectives and hindering AI’s global benefits. The study emphasizes that AI models are influenced by subjective perspectives at every development stage, from data collection to model alignment, primarily reflecting Western viewpoints.
The researchers call for a reevaluation of data collection practices to include broader demographic and cultural perspectives. This inclusion can significantly enhance model performance and ensure AI serves a more diverse audience. Engaging diverse cultures during the model alignment process and combining human evaluations with automatic metrics can improve AI’s real-world performance, especially in non-Western communities.
Funding sources also shape AI development priorities. Nations without significant economic incentives for AI research might prioritize short-term gains, potentially stifling long-term AI benefits. Philanthropic initiatives and government support are essential to encourage inclusive AI development.
The study concludes by stressing the need for input from a wide range of perspectives to develop AI systems that benefit everyone. Incorporating diverse viewpoints, reassessing data practices, and expanding evaluation metrics will ensure AI systems are more inclusive and effective globally.
