An investigation by Alex Reisner for The Atlantic found four music datasets circulating among AI developers that together hold more than 21 million recordings, much of it copyrighted. They mix stars such as Taylor Swift, Bad Bunny, Billie Eilish and the Beatles with tens of thousands of little-known independent acts, plus jazz and classical material.
Two of the collections hold a little over 100,000 tracks each; the other two contain roughly 9 million and 12 million. Each has been downloaded thousands of times, but because developers rarely disclose training data, it is mostly unknown who used them. The report says Google and Stability AI drew on one of the smaller sets, the Free Music Archive.
Neither dataset with a known origin was built for commercial music generators. The largest, LAION-DISCO-12M, was published in 2024 by the German non-profit LAION for academic research; it lists links to public YouTube tracks rather than the audio, and LAION warns against commercial use. The Free Music Archive set was assembled by researchers in 2017 from a library of Creative Commons tracks run by the radio station WFMU, licences written long before generative AI.
The findings arrive with Suno and Udio facing at least a dozen lawsuits by The Atlantic's count. Universal and Warner have moved from suing to licensing, while Sony is still litigating against both companies, GEMA and Koda are suing Suno, and independent artists and the American Federation of Musicians have brought cases of their own. One act, The American Dollar, said Suno had cut its licensing income by almost 80%.
The wider cost is hard to pin down. A CISAC-commissioned study projected that generative AI could take nearly a quarter of music creators' revenue by 2028, and Deezer said in April that close to 75,000 fully AI-made tracks were arriving every day, over 44% of its new uploads, though they drew only a small share of listening.
Sources · 1