Artificial intelligence trained to generate music overwhelmingly favors Western genres, leaving non-Western musical traditions severely underrepresented and poorly reproduced, according to new research published in 2025.
A study from Mohamed bin Zayed University of Artificial Intelligence found that 94% of training data used in AI music models comes from Western musical traditions. African music accounts for just 0.3% of the datasets, Middle Eastern music 0.4%, and South Asian styles 0.9%. The research team analyzed music generation across 147 languages and styles from 79 countries using a dataset called GlobalDISCO.
When researchers asked AI models to generate music in regional styles, the systems often defaulted to outputs closer to mainstream Western sounds rather than authentic regional music. A model prompted to create Turkish makam or Hindustani classical pieces produced results that lacked the tonal complexity and rhythmic patterns central to those traditions.
The MBZUAI team tested how well models performed on non-Western music, finding that simply fine-tuning existing models did not solve the problem. \”The results suggest modifying models is not enough, addressing dataset bias is essential to building inclusive music generation systems,\” the researchers wrote.
AI models perform better on music from high-resource, mainstream regions and genres, producing lower-quality or less authentic outputs for niche or lower-resource local styles. The bias originates in how training datasets are assembled. Most AI music systems learn from commercially available recordings, which disproportionately represent Western pop, rock, classical, and electronic music.
The Responsible AI International Community, an initiative launched in 2024, notes that AI’s rapid development outpaces artists’ ability to adapt creatively. The group warns this could lead to perceptions of low artistic quality and loss of cultural nuance in AI-generated art and music.
Researchers and musicians caution that the bias risks narrowing the global musical landscape. If AI systems continue to reinforce dominant Western aesthetics, they could alienate non-Western audiences and contribute to cultural homogenization. Platforms using AI for music curation or generation might inadvertently suppress regional styles that algorithms struggle to recognize or reproduce accurately.
Separate research published in PLOS One examined whether AI-generated music produces the same emotional responses as human compositions. The study involved 88 participants who listened to audiovisual pieces, some scored with AI music and others with human compositions.
The GlobalDISCO dataset has been released as a public resource to help researchers identify and address biases in music generation systems. The tool provides benchmarks for evaluating how well AI models handle diverse musical traditions, offering a framework for measuring improvement as developers work to rebalance training data.
