Here’s Everything Going On in AI and Music Right Now

A month ago, The Atlantic published an investigation into four publicly circulating datasets that together contained roughly 21.2 million copyrighted recordings. Accompanying it was a free tool that allowed artists to check whether their songs appeared in the collections. In the week that followed, artists whose music appeared in the results expressed their dismay, with some wondering what legal recourse they had.

A nuance that got lost in the noise is that a song's presence in a dataset is not proof that any specific company trained on it. Three of the four datasets contain no audio at all, only links to YouTube and Spotify paired with metadata. Developers typically feed these into automated ripping tools that bypass logins, adverts, and the mechanisms that pay creators. The only confirmed users are Google and Stability AI, which cited the Free Music Archive, the smallest of the four collections, in published research papers. Nobody knows who downloaded what from the two datasets containing the overwhelming majority of the 21 million tracks because there is currently no legislation requiring disclosure.

Amid the fallout, Sleeping-DISCO-9M, one of the four datasets, comprising just over 9.7 million tracks scraped from commercial music, was taken down by its publisher, the Sleeping AI Research Collective. Removing the original download link is no guarantee that the data is no longer circulating, and the dataset remains searchable through The Atlantic's tool.

A paper accompanying the dataset was also quietly withdrawn. Titled SLEEPING-DISCO 9M: A Large-Scale Pre-Training Dataset for Generative Music Modeling, it described the dataset as one that reflected the music of the real world: "there are no open-source high-quality datasets representing popular and well-known songs," the abstract read.

The authors intended the dataset for pre-training generative models, and while removing it is not an admission of liability, plaintiffs could argue that it demonstrates awareness of potential copyright concerns. Without it, researchers hoping to build open-source music models are pushed back towards royalty-free libraries, Creative Commons music, licensed catalogs, or synthetic datasets, all of which come with limitations in scale or musical diversity.

Sleeping AI has since issued a statement on its website saying that the "absolute erasure" was a deliberate decision by the team "in response to emerging public controversies, localized targeting behaviors, and escalating incidents of doxing that compromised the personal security, privacy, and individual liberties of independent authors."

The statement goes on to deny that the dataset hosted, indexed, or distributed copyrighted multimedia content, adding that "nor were any generative AI models ever trained on this metadata."

The Legal and Ethical Battle Over AI Data Extraction

While the appearance of a song in The Atlantic's searchable datasets has not resulted in legal action by affected artists, a separate wave of class action lawsuits filed by independent musicians is now making its way through U.S. courts, targeting AI music companies Suno, Udio, Mureka and, more recently, Google over allegations of unauthorized copying and AI training. "If independent voices stay silent, AI companies will lock in a business model that treats creative work as free fuel. If we act now, together, we can force them to license, pay, and respect the artists who make music worth hearing," reads an article on civil rights law firm Loevy + Loevy's website.

Back in April, we reported that licensing talks between Suno and two of the three major labels had collapsed. Three months on, the fracture has hardened into a fault line over a single, pressing question: whether AI-generated songs can be downloaded from the platforms that make them. Warner, which settled with Suno in November, allows paid users to download and commercialize their creations. Universal, which settled with Udio in October, required Udio to disable all downloads and gave users a 48-hour window to export their songs before the walls went up.

Universal and Sony have meanwhile gone to court to see the terms of Warner's deal, arguing that they cannot negotiate blind against an agreement whose contents they are not allowed to know. Sony, which refuses to settle, added more than 61,000 recordings to its case against Suno. That case is headed for a summary judgment hearing this month. The ruling may provide answers to whether training AI models on copyrighted recordings without a license is fair use or theft.

A paper published in April in the Eastern African Journal of Humanities and Social Sciences argues that AI training datasets are “political artifacts that fix neo-colonial forms of oppression by freezing dynamic, living cultures to be consumed exogenously.” It traces back to the analog blueprint: Solomon Linda's 1939 recording of “Mbube” was taken from its creator, circulated globally, and monetized by everyone but him. That a dataset called Sleeping-DISCO sits at the center of the current controversy is an irony Linda's estate did not need.

Legal scholars Chijioke Okorie and Melissa Omino have made a related argument about the licensing frameworks now being proposed as the solution, noting that standard open licenses treat African dataset owners as though they were well-resourced Global North entities. They argue that the result is an inequity born of treating unequally situated actors as equals. 

Some builders on the continent, like Nigerian start-up Korin AI, are responding by rejecting the scraping model altogether. It works by licensing music from local production firms and paying singers to record training data. "Those guys abroad, they just scrape the internet for data and don't pay people. We don't want to do that," founder Oluwaseun Olajide-Philips told OkayAfrica in April.

Olajide-Philips's refusal challenges the assumption held by many commentators that extraction is the inevitable foundation of AI development; that we should all just get with the program. 

Johannesburg-based Lelapa AI's Pelonomi Moiloa, writing in Nature, argues that the scaling model underpinning contemporary generative AI was developed under “conditions of extraordinary abundance”: cheap capital, plentiful energy, vast computing infrastructure and seemingly limitless data. Moiloa looks past the hallucinated citations of South Africa's since-withdrawn draft national AI policy and points at the assumption that the extractive Silicon Valley blueprint is the only path forward. Her company develops language AI systems designed for African languages and other resource-constrained environments.

The regions with the greatest linguistic diversity, of which the African continent is among, may be the first to expose the structural limits of a model built on this scraped abundance. The takedowns, the lawsuits, the pending summaries about the legal limits of unauthorized AI model training — they all revolve around the question of who pays for the data. Moiloa's question is simple: does the machine really require 21 million songs to learn how to sing?