Music AI Models Are Amplifying Decades of Industry Bias Against Non-Western Genres

Researchers at Abu Dhabi's Mohamed bin Zayed University of Artificial Intelligence have found that Western genres account for 94% of the training data behind leading generative music tools, while music from Africa (0.3%), the Middle East (0.4%), and South Asia (0.9%) is severely underrepresented—despite those regions comprising roughly half the world's population. The findings, presented at the 2025 NAACL conference, showed that when models attempted to generate an Indian raga or Turkish Makam, they defaulted to Western tonal structures, and that feeding the models additional corrective recordings actually degraded output quality—the Western training data was too dominant to override.

The study argues that this is not simply a data-volume problem but a structural one: decades of music-industry decisions about who gets signed, tracked, and promoted are now embedded in AI training sets that will shape recommendation, discovery, and monetization for years to come. The bias extends to gender—women represented just 14.5% of Billboard Hot 100 songwriters and 4.4% of producers in 2025, figures that have barely moved since 2012 according to the USC Annenberg Inclusion Initiative. A 2024 MediaFutures/University of Bergen survey separately confirmed that popularity bias is among the most persistent forms of algorithmic unfairness in recommendation systems, while a 2026 AFEM survey found that half of 22 music-tech companies cite conflicting metadata as their single biggest structural challenge.