The Atlantic has identified four music datasets containing tens of millions of songs that AI developers have used or circulated for model training, putting hard names and numbers on a copyright fight that has already moved from theory to court.
Alex Reisner's investigation for The Atlantic, published June 15, gives the AI music industry the thing it has spent two years trying not to provide: a searchable trail of training data. The report identifies four datasets, including a large collection assembled by the German AI non-profit LAION with roughly 12 million tracks, a second with about 9 million, and two smaller sets of around 100,000 songs each. According to The Atlantic, those databases include hundreds of tracks apiece from Taylor Swift, ABBA, the Beatles, Snoop Dogg, and Michael Jackson.
The force of the story is not that artists suspected this. Many have. The force is that the databases are searchable. An artist no longer has to argue from instinct or from a strange imitation buried inside a generated song. They can look up a name and see whether a catalogue appears in a dataset that developers used or passed around. That is a different kind of evidence, and it is harder for the industry to wave away.
The connection to Google and Stability AI is also important, though it should be stated carefully. The Atlantic tied both companies to the Free Music Archive dataset, one of the smaller collections it examined. That does not mean every company named in the reporting trained the same kind of commercial music generator on the same material. It does mean the provenance problem is no longer confined to small startups testing the edge of copyright law. Large AI companies have been close enough to these datasets that the question now lands in boardrooms, not only in Discord channels and research repositories.
The legal fight was already moving before The Atlantic's database landed. The Recording Industry Association of America sued Suno and Udio in June 2024 on behalf of Universal Music Group, Sony Music Entertainment, and Warner Music, alleging that the companies used copyrighted recordings without permission to train music generation models. Wired reported at the time that the cases were filed in Massachusetts against Suno and in New York against Udio, with damages sought of up to $150,000 per infringed work.
Suno and Udio have not treated the allegation as a simple denial. In 2024 filings covered by The Verge, both companies argued that training on copyrighted material could be protected as fair use. That is the issue the music business has been waiting to test. The labels say the companies copied protected recordings at scale and built products that compete with the works they absorbed. The AI companies say model training is a transformative use that helps people make new songs. The Atlantic's reporting does not settle that legal argument, but it gives the plaintiffs a more concrete map of what was allegedly in the pipeline.
Some of the fight has already moved from lawsuits to licensing. The Associated Press reported that Universal Music Group settled with Udio in October 2025 and agreed to work with the company on a new controlled music creation and streaming platform. Warner Music Group later settled with Suno and announced a licensing partnership, with Warner's former Songkick business also moving to Suno as part of the arrangement. Sony has not resolved the same core fight in the same way, which is why the remaining litigation still matters.
The problem is bigger than Suno and Udio
The immediate risk is not only for the companies that built the models. It is for every startup that built a product on top of them and every investor who assumed the licensing issue would be handled somewhere else. If a model provider's training data becomes a live legal problem, the next questions are predictable: who used the model, who sold access to outputs, who marketed the product as commercially safe, and who performed diligence before writing a check.
That is why data provenance has become a real diligence item, not a box at the end of a legal memo. A music or media AI startup now has to answer plain questions. What recordings were used? Who owns them? What licenses exist? Which datasets were excluded? Those questions used to sound like cautious lawyer talk. After The Atlantic made the datasets searchable, they sound like the minimum any serious buyer, partner, or investor should ask.
The timing also tightened because the music industry has opened a second front. Pitchfork reported last week that the American Federation of Musicians sued Universal and Warner, alleging the labels licensed recordings for AI uses without properly compensating session musicians under their collective bargaining agreements. That case is not the same as the RIAA's 2024 suits against Suno and Udio, but it points to the same pressure point. Even when labels and AI companies strike licensing deals, the money and permissions may not be settled all the way down to the people who played on the recordings.
The code copyright fight around GitHub Copilot showed how long these cases can run. The Joseph Saveri Law Firm filed that class action in 2022 against GitHub, Microsoft, and OpenAI, arguing that public code was used without respecting license terms. Music is traveling a related road, but with a cleaner plaintiff base, a richer licensing history, and recordings that are easier for ordinary listeners to understand than open-source code fragments.
The hard part for AI music companies is that a good product does not erase a bad data trail. Suno and Udio can argue fair use. They can sign licensing deals. They can move future models toward authorized catalogues. But the old datasets still exist, and now at least some of them can be searched by the artists whose work appears inside them.
Also read: The federal government is treating xAI's data center as a national security asset and that should concern everyone • Kingboard Laminates' 148% stock surge shows where the real AI infrastructure money is flowing • Morgan Stanley opens its trillion-dollar stock plan plumbing to AI agents and Wall Street is watching
