Millions of copyrighted songs were used to train AI music models, according to evidence assembled by The Atlantic. The material includes contemporary global hits, and in some cases, the rights holders never agreed to have their work fed into model training. That turns a long-running suspicion into something much harder for AI firms to shrug off.
The reporting matters because the fight is no longer abstract. Suno has acknowledged in court filings that it trained on tens of millions of recordings, some protected by copyright, while music companies are now treating the issue as a licensing dispute with real price tags attached. In the background, the broader market is already shifting from ”train first, ask later” to a more expensive model built around permissions, contracts, and damage control.
What The Atlantic’s catalogs show
Journalist Alex Reisner published four catalogs identifying music used in model training, giving the argument a level of specificity that AI companies have tried to avoid. The list reportedly includes artists such as Taylor Swift and Bad Bunny, both of whom were not asked to let their work be used in that way.
- Suno said in legal documents that it trained on tens of millions of recordings.
- Sony, UMG, and Warner have sued Suno and Udio.
- The labels are seeking up to $150,000 for each track that ended up in a training set.
Fair use is doing a lot of heavy lifting
AI developers still argue that training falls under fair use because models absorb patterns rather than store songs. Rights holders call that framing nonsense, and not quietly: the music industry has described the practice as piracy dressed up for investors. That is blunt language, but the labels have a point – if the output depends on huge unlicensed catalogs, the legal theory starts to look more like a sales pitch than a defense.
The U.S. Copyright Office added another layer of pressure in January 2025, saying AI-generated outputs lack copyright protection when human involvement is insufficient. That leaves a messy gap: the training data may be contested, and the resulting content may also sit outside normal copyright protection. Great news, if your business plan is uncertainty.
AI music licensing deals are already here
There is a reason this battle is increasingly being fought in contracts, not just court filings. Researchers at the University of Tennessee built HarmonyCloak, a tool that adds audio distortions meant to interfere with AI training while staying inaudible to people, but legal agreements may do more damage to the old free-for-all than technical tricks ever will.
Warner Music Group and Universal Music Group have already signed licensing agreements with Udio and Suno, and Tennessee has passed a law aimed at protecting musicians’ voices from unauthorized cloning. Expect more labels to push for the same outcome: if AI companies want the catalog, they will have to pay for it. The open question is how many model makers can afford that bill before the ”creative” music boom becomes a much smaller business.