MBZUAI study reveals bias in AI music models
Study finds that 94% of training data comes from Western music genres
#UAE #culture – A new study by researchers at Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) reveals that Generative AI music systems are overwhelmingly biased toward Western genres, with just 5.7% of training data sourced from non-Western styles. Presented at NAACL 2025 in New Mexico, the research evaluated whether parameter-efficient fine-tuning with adapters could improve performance in underrepresented musical styles such as Hindustani classical and Turkish Makam, offering new insights into the limits of current AI models in music generation.
SO WHAT? – As Generative AI develops greater multimodal capabilities, cultural diversity becomes even more important. This study highlights the extent of the cultural inclusivity gap. Without better global representation in training data, music production AI systems rung the risk reinforcing a narrow creative scope and limiting innovation, and potentially alienating vast global audiences.
Some key points about this MBZUAI study:
Researchers at Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) have found that Generative AI music systems are overwhelmingly biased toward Western genres.
MBZUAI researchers found 94% of generative music model training data comes from Western genres, whilst only 0.3% of data originated from Africa, 0.4% from the Middle East, and 0.9% from South Asia.
The study introduced adapter-based fine-tuning on Hindustani classical and Turkish Makam music using MusicGen and Mustango text-to-music AI models.
Adapters represented just 0.1% of total model parameters, showing a lightweight intervention with targeted goals.
Fine-tuning led to an 8% improvement for Mustango on Hindustani classical and 4% for MusicGen on Turkish Makam.
However, fine-tuning also caused performance degradation in Western genres, revealing a trade-off between specialisation and retention.
Music experts used a modified Bloom’s Taxonomy to assess the musical output, evaluating creativity, analysis, and recall.
The results suggest modifying models is not enough, addressing dataset bias is essential to building inclusive music generation systems.
The Music for All research study was led by Atharva Mehta, a research associate at MBZUAI, with contributions from Shivam Chauhan, Amirbek Djanibekov, Atharva Kulkarni, Gus Xia, and Monojit Choudhury, all of MBZUAI.
Researchers plan to continue working on genre-sensitive evaluation methods and data augmentation strategies for diverse musical cultures.
ZOOM OUT – The UAE has positioned itself at the frontier of AI ethics, inclusivity, and governance, and this music bias study follows on from MBZUAI’s research on cultural and linguistic inclusivity. Earlier this year, the university released groundbreaking All Languages Matter Benchmark (ALM Bench) evaluating large multimodal models across 100 languages and 13 cultural dimensions, revealing significant performance gaps in underrepresented languages and cultural contexts. By publishing open-source tools and empirical evaluations, MBZUAI is helping to set new standards for inclusivity in AI development.
[Written and edited with the assistance of AI]
LINKS
Music for All Research Paper (arXiv)
Read about MBZUAI’s earlier cultural inclusivity research:
New benchmark challenges inclusivity of global AI models (Middle Est AI News)