Africa, the Middle East and South Asia represent roughly half the world’s population, and they are also home to hundreds of distinct musical traditions. But in the training datasets most commonly used to build music AI models, music from Africa accounts for only 0.3%, the Middle East: 0.4%, and South Asia 0.9% — whereas Western genres make up 94%.

These numbers come from researchers at Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence, who surveyed the training datasets behind today’s generative music tools and presented the findings at the 2025 Nations of the Americas Chapter of the Association for Computational Linguistics (NAACAL).

When those models tried generating music in the tradition of an Indian raga, they defaulted to a sitar playing Western tonal structures, producing something that sounded Western with an Indian instrument on top. The same study tested Turkish Makam, a melodic system built on intervals that don’t exist on a Western piano. Once again, the models flattened those intervals into standard Western pitch. When the researchers fed the model additional Hindustani Classical and Turkish Makam recordings to correct the bias, its creative output actually got worse. The Western training data was too dominant to override.

This study confirms that the problem goes deeper than underrepresentation, with the biases embedded in decades of music data now being built into AI systems trained on that data. And it’s these systems that will shape what gets heard, paid and promoted for years to come.

The datasets powering the music industry have been shaped over decades by who’s gotten signed, which markets have been considered worthy of tracking, and which genres received investment. Industry infrastructure is built around particular slices of the business, and it’s treated as though it represents the whole thing. For a long time, those gaps sat quietly in back-office databases, and the consequences were slow. Now those gaps sit in the training data, and the systems built on top of them will run for the foreseeable future.

The bias extends to gender. In 2025, women represented 14.5% of songwriters on the Billboard Hot 100 and only 4.4% of producers, according to the USC Annenberg Inclusion Initiative, which has been tracking these numbers for over a decade. Those figures have barely shifted since 2012. Algorithms are learning from a foundation that doesn’t reflect what people want to hear, but rather reflects what’s already popular, promoted and playlisted at scale. Those outputs are then fed straight back into the loop. A song added to Spotify’s Today’s Top Hits generates millions of streams, which tells the algorithm it’s popular, which feeds more recommendations, which generates more streams.

For independent artists, emerging scenes and non-Western music, the cycle works the other way: less data, less visibility, fewer recommendations. A 2024 survey published by MediaFutures and the University of Bergen, Norway, confirmed that popularity bias is one of the most persistent and well-documented forms of algorithmic unfairness in recommendation systems.

None of this started with AI. The metadata problems have been building for years and cluster in foreseeable areas: around independent artists, non-Western catalogs and anything released outside the major label pipeline. In early 2026, the Association For Electronic Music (AFEM) surveyed 22 music tech companies. Half identified conflicting metadata across databases as their single biggest structural challenge, and 41% pointed to the absence of universal artist and song identifiers.

The industry has a deeply embedded habit of releasing music first and sorting out the data later. A track travels from artist to distributor to DSP to collection society, and at every handoff, something can go wrong. Broken metadata means broken everything downstream: wrong recommendations, wrong royalties, wrong training data. The Global Repertoire Database attempted to fix this in 2014 but failed, as the performing rights organizations couldn’t agree on governance. Institutions protect what gives them power, and that tension hasn’t gone away.

What has changed is the speed of the inequity. Deezer says it now receives roughly 75,000 fully AI-generated tracks every day. As of April 2026, that accounts for 44% of all new uploads to the platform (for reference, in January 2025, the streaming service claimed it was only 10%; by November, it said it had jumped to 34%.) A joint study by Deezer and Ipsos in late 2025 found that 97% of listeners couldn’t tell whether a track was AI-generated or human-made.

Now, it’s not quality but sheer volume that filters music. Spotify reported removing more than 75 million spammy tracks in the 12 months leading up to September 2025, and Deezer said it found that 85% of streams on AI-generated tracks were fraudulent. The platforms are firefighting. When upload numbers double in less than a year, recommendation systems respond the only way they can: by leaning harder on existing play counts, saves and playlist placement history.

As such, tracks with momentum get more of it, while those without it disappear faster. This is how bias gets worse.

The volume of AI-generated music doesn’t land evenly, either. Established artists have cultural authority and loyal audiences. The artists getting squeezed are the ones in the middle: producers with a small fanbase, indie-label acts who’ve been building a community for years. They’re pressed from below by a rising floor of technically passable artificial music and from above by recommendation systems that reward whatever already has traction. UK Music’s 2025 creator survey captured the mood among creators: 66% said AI threatens their careers, 92% want labeling and 93% want AI companies to pay when they use compositions for training.

There’s also a discovery problem that doesn’t get discussed enough. Chartmetric, the most widely used A&R scouting platform, tracks more than 12 million artist profiles. Warner Music’s Sodatone, Soundcharts and similar tools all draw from the same pools of streaming and social data. If every A&R team is running the same algorithmic scouting tools on overlapping datasets, any artist that those tools bring to the surface has, by definition, already generated enough data to be “discovered”. Meanwhile, the genuinely early talent, the artists recording in bedrooms or community studios, exists in spaces where these platforms have no reach.

Financial consequences of this broken system can be put into context: Fair Play, an independent audit of U,K. nightclub royalties published in November 2025, looked at how performance royalties get calculated for electronic music. The method that carries the most weight extrapolates from monitoring data drawn almost entirely from commercial clubs or radio plays. Underground venues, independent producers and music that travels outside the mainstream — all of it gets measured against a sample that was never designed to include it. According to the audit, only 36% of electronic music performances in UK nightclubs result in the correct creator being paid. That’s an estimated 5.7 million pounds going to the wrong people every year, in nightclubs alone.

Even though the finding is specific to the U.K. and to electronic music, the logic still holds. Wherever a system relies on monitoring a subset of the market and applying those findings to the whole, the sample becomes the standard, and everyone outside it ends up subsidizing everyone inside.

These challenges also directly affect which products get developed. The same AFEM survey found that 29% of music tech companies pivoted away from their original product vision because the data and rights infrastructure made the original idea impossible. Half of respondents said that more than 50% of their engineering resources go towards workarounds rather than building the actual product.

Bias in music data is not new. Now, though, it is being built into AI systems that have been trained on historical datasets that encode every gap the industry never fixed. Some of the infrastructure is starting to catch up. The EU AI Act now requires AI developers to disclose what copyrighted material they trained on. In April 2026, Spotify launched a beta built on DDEX, the industry’s metadata standard body, allowing artists to disclose how AI is used in their music. Similarly, Apple Music, Bandcamp and Deezer are enforcing AI disclosures in their own way. Universal Music has also begun embedding the International Standard Name Identifier (ISNI) across its catalogue. These measures address transparency and attribution, not the underlying imbalance in the data itself. But disclosure is where accountability starts.

We are seeing solutions emerge, and there is progress, but is it happening fast enough? The industry has lived with this infrastructure problem for decades without fixing it. AI has turned finding this solution into a deadline.

Rufy Ghazi is a music business professional with experience in product management, digital transformation and research. She works as a product consultant for music tech companies, drawing on experience at ByteDance (TikTok), Amra (Kobalt Music Group) and various early-stage startups. She is the Head of Music Research at Audience Strategies, where she spearheads data-driven research projects. Her notable reports include the UK Electronic Music Industry Report (for NTIA), Sound Investments (MTUK’s study of the UK music tech ecosystem), and A Slice of Fairness (for Aslice). She has also led the Fair Play initiative, an independent audit of rights and royalties in electronic music.

Daily newsletters straight to your inbox

Sign Up