The Atlantic's databases let artists, labels, and plaintiffs check whether specific tracks were used to train models like Suno and Udio, as the lawsuits against the AI music generators move through the courts.
The Atlantic has turned a sprawling copyright fight into something a plaintiff can search.
On Monday, the magazine published four searchable databases of music used to train AI music generators, totaling roughly 21 million tracks, including catalog hits from Taylor Swift, Bad Bunny, and other recognizable artists, according to Engadget's summary of the investigation. The release gives artists, labels, and lawyers a public lookup tool to test whether a specific song ended up in a training set. That is the kind of evidence that has been hard to pin down in the lawsuits already moving through US courts.
The numbers are large enough to matter on their own. The Atlantic's reporter Alex Reisner, writing for the magazine's AI Watchdog project, says one database contains about 12 million tracks and another about 9 million, with two further databases each holding around 100,000 songs, according to the underlying article. The publication of those databases is the news peg: the rest of the copyright story now has a public, citable index.
The litigation is already in motion. The Recording Industry Association of America sued two of the most prominent AI music generators, Suno and Udio, accusing them of wholesale scraping of copyrighted recordings, and the defendants have leaned on fair-use defenses, per Engadget's prior coverage of the suits. The Atlantic's investigation lands as those cases progress and as plaintiffs in other AI copyright fights look for templates.
The closest analog comes from the book world. A federal judge rejected a record-breaking $1.5 billion settlement in a copyright class action against the AI company Anthropic brought by authors, according to Engadget. The deal had been reached after piracy claims proved more legally durable than broader copyright arguments, and the rejection leaves payout terms unresolved. Music plaintiffs are watching that case closely because the new music databases are a different kind of evidence: a public, searchable record rather than a sealed internal set.
The platforms named in the lawsuits are not the only ones grappling with the fallout. Spotify is testing tools to surface and manage AI-generated tracks, often called "AI slop" by the industry, on artist profiles, according to Engadget. Deezer is rolling out an AI music detector, and Apple Music is offering distributors the option to flag AI content. The enforcement gap was visible in a recent case in which an AI-generated copycat of the band King Gizzard and the Lizard Wizard sat on Spotify for weeks before being noticed, per Engadget's reporting.
The scale The Atlantic reports carries a caveat. The published database sizes, roughly 12 million, 9 million, and two at around 100,000, are figures the magazine assembled, not audited counts from the AI companies themselves. The fact that a track appears in a database is also not, on its own, proof of infringement: the legal questions around fair use, licensing, and the line between training data and copied output are exactly what the Suno and Udio cases are now testing.
What changes on Monday is the evidence economy. An artist who suspects a song was used to train a model can now look it up, and a label considering a claim has a public starting point. Whether that is enough to shift the legal track remains the open question, and the next substantive filing in the Suno and Udio cases will be the first real test.