Sleeping-DISCO-9M Dataset Removed, Sparking Class-Action Lawsuits Against AI Music Companies

The Atlantic’s June 12 exposé on four sprawling music datasets—together holding roughly 21.2 million copyrighted tracks—has forced the largest of them, Sleeping‑DISCO‑9M, to vanish from public view. Published by the Sleeping AI Research Collective, the 9.7‑million‑track collection was taken down a month after the investigation, though the data can still be queried through the Atlantic’s free AI‑Watchdog interface.

The Atlantic’s investigation catalogued four openly accessible datasets. The first, a 12‑million‑track set, remains largely a catalogue of YouTube and Spotify URLs paired with metadata, with no audio files. Sleeping‑DISCO‑9M, the 9.7‑million‑track trove scraped from commercial services, was the only dataset that actually contained audio. The remaining two were the 106‑k‑track Free Music Archive set and the 114‑k‑track Spotify‑Tracks set. In practice, only Google and Stability AI have publicly cited the Free Music Archive in their research; there is no evidence that any AI firm has trained a model directly on Sleeping‑DISCO‑9M.

After the removal, the paper that had accompanied the dataset—"SLEEPING‑DISCO 9M: A Large‑Scale Pre‑Training Dataset for Generative Music Modeling"—was quietly withdrawn. The article had described the collection as a realistic representation of popular music and lamented the lack of open‑source high‑quality datasets for well‑known songs. The authors claimed the data was intended for pre‑training generative models, but the withdrawal does not equate to an admission of wrongdoing. In a statement on its website, Sleeping AI explained that the “absolute erasure” was a calculated response to public backlash, targeted behaviors, and doxing incidents that endangered independent artists’ privacy.

The takedown has amplified a wave of class‑action lawsuits filed by independent musicians in U.S. courts. The suits name Suno, Udio, Mureka, and Google, alleging that these companies copied copyrighted recordings without permission for AI training. The claims invoke the Copyright Act of 1976, which treats sound recordings as a distinct category of copyrightable work. While no artist has yet pursued legal action based on the Atlantic tool’s search results, the lawsuits aim to force AI firms to license music, pay royalties, and honor artists’ rights.

Licensing negotiations have also grown more fraught. Suno’s November 2025 settlement with Warner Music Group permits paid users to download and commercialise AI‑generated songs, whereas Universal’s October 2025 settlement with Udio required the platform to disable downloads and provide a 48‑hour export window. Universal and Sony have sued over Warner’s agreement, arguing that they cannot negotiate in good faith while blind to its terms. Sony added more than 61 000 recordings to its case against Suno, and a summary‑judgment hearing is scheduled for this month. The ruling could clarify whether training AI models on copyrighted recordings without a license falls under fair use or constitutes infringement.

The controversy has also prompted reactions from African AI developers. Nigerian startup Korin AI rejects the scraping model, instead licensing music from local production firms and compensating singers directly. Johannesburg‑based Lelapa AI’s Pelonomi Moiloa, writing in Nature, argues that the scaling model underlying contemporary generative AI relies on an “extraordinary abundance” of data and resources that many regions lack. Moiloa questions whether a machine truly needs 21 million songs to learn how to sing, highlighting structural inequities in current licensing frameworks.

The takedown of Sleeping‑DISCO‑9M, the pending lawsuits, and the push for alternative data collection methods underscore a broader industry debate over who should pay for the data that fuels generative AI. As courts weigh the legality of training on copyrighted recordings, the music community remains split between those demanding licensing and payment and those advocating for open‑source or royalty‑free alternatives.

The Atlantic’s AI‑Watchdog tool continues to provide a searchable interface for the four datasets, allowing artists to verify whether their work appears in any of the collections. The ongoing legal and industry developments suggest that the question of data provenance will remain a central issue for AI music companies and creators alike.