Journalists at The Atlantic say they have dug up hard evidence that millions of copyrighted songs were used to train AI music models, including recent global hits whose rights holders never agreed to that use. The dispute is no longer about vague fears of scraping; it is about named catalogs, named services, and a legal fight that is already pulling in major labels, lawmakers, and the U.S. Copyright Office.
The reporting centers on four catalogs published by Alex Reisner that map specific music used in training sets. In court filings, Suno has acknowledged training on tens of millions of recordings, some of them protected by copyright and used without permission from artists and rights holders, including Taylor Swift and Bad Bunny. That is the kind of detail companies hate because it turns an abstract AI debate into a spreadsheet.
Major labels are already suing
Sony, UMG, and Warner have filed lawsuits against Suno and Udio, seeking up to $150,000 for each composition that ended up in a training set. The AI companies argue that training falls under fair use and that models learn broad patterns rather than memorizing songs. The music industry’s answer is simpler and sharper: if a system was built on unlicensed recordings, calling it innovation does not make the invoices disappear.
- Suno has said it trained on tens of millions of recordings.
- The catalogs published by Alex Reisner identify specific music used for training.
- Rights holders named in the dispute include Taylor Swift and Bad Bunny.
- The labels are asking for up to $150,000 per song.
The legal pressure on AI music is getting heavier
The US Copyright Office added another problem in January 2025, saying outputs from such systems are not protected by copyright because they lack sufficient human involvement. That leaves AI-generated music in a strange place: easy to produce, hard to defend, and potentially expensive to commercialize if the training data was dirty to begin with. The old ”we only learned the style” argument suddenly sounds less clever when the paperwork is this specific.
Researchers at the University of Tennessee have also built HarmonyCloak, a tool that adds inaudible distortions to audio files to make them harder for AI systems to learn from. It is a technical countermeasure, but the battlefield is shifting fast toward licensing deals instead of purely defensive tools. Warner Music Group and Universal Music Group have already reached licensing agreements with Udio and Suno, and Tennessee has passed a law protecting musicians’ voices from improper cloning.
What happens next for AI music
The direction is fairly obvious: the biggest players are being pushed toward paying for access, while smaller startups will struggle to defend training on copyrighted catalogs at scale. That does not end the fight, because the next clash will be over how much a license should cost and whether retroactive settlement becomes the price of entry. For now, the industry has moved from denial to damage control, and the labels are making sure everyone can read the receipts.