The Atlantic Publishes Databases Showing 21 Million Songs Were Used to Train AI Without Artist Consent

What The Atlantic Found

On June 15, 2026, The Atlantic published four searchable databases compiled by staff writer Alex Reisner documenting music used to train AI models. The numbers are large: one database contains 12 million tracks, a second contains 9 million, and the remaining two hold roughly 100,000 songs each. That's more than 21 million tracks total.

The artists represented include Taylor Swift and Bad Bunny — household names with active legal teams and recording contracts. Their presence in these datasets is not alleged. It's documented and searchable.

The Legal Context

Lawsuits against AI music platforms Suno and Udio are already underway, according to Engadget's June 15 coverage of the story. Both companies have defended scraping copyrighted recordings under a fair use argument, a legal theory that has had mixed results in court.

A parallel case in book publishing — where authors sued an AI company over wholesale scraping of their works — did not succeed on copyright infringement grounds, but piracy allegations proved more durable. According to Engadget, the initial settlement in that case was $1.5 billion, with the full payout still pending as of June 16, 2026.

The music industry is watching that outcome closely. The Atlantic's databases give copyright holders something they previously lacked: a searchable, specific record of whose work was used and in what volume.

What AI Companies Would Say

The fair use defense is not frivolous. Under U.S. copyright law, transformative use of copyrighted material for purposes like commentary, research, or creating a fundamentally new product can be legally protected. AI companies have argued that training a model on music is different from copying and distributing songs, and that the output — generated audio — is sufficiently distinct from the training input.

That argument has not been fully adjudicated in the music context. Courts are still working through what fair use means for large-scale AI training, and legal scholars are genuinely divided. Dismissing the defense outright would misrepresent where the law actually stands right now.

What Artists Are Dealing With

Separate from the lawsuits, music streaming platforms have been scrambling to contain the downstream effects of AI-generated music. Engadget reported that services have taken steps to prevent, identify, or label AI-generated tracks, though with varying success.

Scammers have been creating AI imitations of existing bands and monetizing them. They're profiting from the sound and brand of artists who never consented. This is a problem that exists regardless of how copyright litigation ultimately resolves. It's happening now, and the platforms are not fully on top of it.

Source Note

The Atlantic's article by Reisner is the primary investigative work here, and the databases it published are the core of the story. Engadget's coverage accurately summarized the scope. One gap in both sources: neither contacted Suno, Udio, or any AI music company for an on-record response to the specific databases published. Their legal positions are referenced only through prior court filings. That matters because the databases' methodology — how tracks were identified as training data — has not yet been subjected to public scrutiny or rebuttal from the companies named.

What Happens Next

The Recording Industry Association of America (RIAA) and individual labels now have detailed, searchable documentation to work with. Whether that translates into new litigation, expanded existing suits, or settlement pressure on Suno and Udio depends on how attorneys assess the databases' evidentiary value — something no source has addressed on the record yet.

The more concrete near-term question is whether the book-publishing settlement's piracy theory travels into music law. The $1.5 billion figure in that case suggests the legal exposure for AI companies could be significant if music plaintiffs can make the same argument stick. Courts have not ruled on that yet, and the full payout in the publishing case remains unresolved as of today.